Gerd Gigerenzer, Jeff Schall, Tom Griffiths, Antonio …€¦ · · 2017-12-15Gerd Gigerenzer,...
Transcript of Gerd Gigerenzer, Jeff Schall, Tom Griffiths, Antonio …€¦ · · 2017-12-15Gerd Gigerenzer,...
DMB 2014 | 912 Sept 2014 | 1
If found, please return to the personabove or to the reception desk.
Gerd Gigerenzer, Jeff Schall,Tom Griffiths, Antonio Rangel
EricJan Wagenmakers,Nick Chater, Daniel Wolpert?
2 | DMB 2014 | 912 Sept 2014
This conference brings together researchers from a wide range ofdisciplines who have an interest in decision making. We think this is aparticularly exciting and important time for research in this area: a timewhere there is real potential for a convergence between theory,experimental work, applied research and policy to affect everydaydecisions that impact on heath and well being. We hope that you willtake this opportunity to hear about new discoveries and perspectivesfrom across the wide range of disciplines represented and hope thatthis will inspire your own research.
The conference is part funded by a grant from the Engineering andPhysical Sciences Research Council (UK) who fund the Decisionmakinggroup here in Bristol University(http://www.bris.ac.uk/decisionsresearch/).
We really hope that you enjoy the conference and find a little time toexplore our beautiful city. Our organising team and volunteers are onhand to help you you can spot us in the red tshirts.
On behalf of the organising team and the members of the Decisionmaking group here in Bristol – welcome!
Iain Gilchrist and David LeslieCoDirectors Bristol DecisionMaking Group
Organising Team:Roland Baddeley, Rafal Bogacz, Rosie Clark, Adnane EzZizi,
Simon Farrell, Rebecca Floyd, Andreas Jarvstad, Casimir Ludwig,Gaurav Malhotra, Alice Mason, John McNamara and SallyAnn Parr y
?
4 | DMB 2014 | 912 Sept 2014
MARLÈNE ABADIE
LUIGI ACERBI
SILVIO ALDROVANDI
FLORIAN ARTINGER
ROLAND BADDELEY
TIMOTHY BALLARD
DAN BANG
ROBERT BEIGLER
ULRIK BEIERHOLM
SHIRLY BERENDS
SUNDEEP BHATIA
ROBERT BIEGLER
VALERIO BISCIONE
WOUTER BOEKEL
RAFAL BOGACZ
ALINE BOMPASS
JEFFREY BOWERS
NICOLA BOWN
NEIL BRAMLEY
DANIEL BRAUN
JULIA BRETTSCHNEIDER
GORDON BROWN
JOANNA BURCHBROWN
EGLE BUTT
NICK CHATER
ROSIE CLARK
MATTHEW COLLETT
JOHN COOK
ANTHONY CRUICKSHANK
GILLES DE HOLLANDER
BARNEY DOBSON
CAROLINA DORAN
JACK DOWIE
GILLES DUTILH
ULLRICH ECKER
TOBIAS ENDRESS
ADNANE EZZIZI
RICHARD FAIRCHILD
A. ALDO FAISAL
GEORGE FARMER
TIM FAWCETT
CAROLINA FEHER DA SILVA
REBECCA FLOYD
KERSTIN GIDLÖF
GERD GIGERENZER
IAIN GILCHRIST
RUSSELL GOLMAN
JOHN GOUNTAS
TOM GRIFFITHS
JOHN GROGAN
RONG GUO
CLAIRE HALES
STEPHEN HEAP
CRAIG HEDGE
ANDREW HIGGINSON
THOMAS HILLS
JANINE HOFFART
JANINA HOFFMANN
AUDREY HOUILLON
ANDREW HOWES
ALASDAIR HOUSTON
LAURENCE HUNT
TOM JAHANSPRICE
ANDREA ISONI
PERKE JACOBS
ANDREAS JARVSTAD
HENRY JORDAN
HANNAH JULLIENNE
DAVID KELSEY
MAX C. KEUKEN
METTE KJER KALTOFT
TOMAŽ KOLAR
TATIANA KORNIENKO
NATHAN LEPORA
DAVID LESLIE
STEPHAN LEWANDOWSKY
KEVIN LLOYD
BRADLEY LOVE
SHENGHUA LUAN
ELLIOT LUDVIG
CASIMIR LUDWIG
JIEYU LVJASMIN MAHMOODI
NISHANTHA MALALASEKERA
GAURAV MALHOTRA
DAVE MALLPRESS
ENCARNI MARCOS
ALICE MASON
ANJALI MAZUMDER
JOHN MCNAMARA
EUGENE MCSORLEY
LIANNE MEAH
SOPHIE MEAKIN
GEOFFREY MÉGARDON
MICHAEL MENDL
JOLYON MILESWILSON
RANI MORAN
TIMOTHY MULLETT
JERYL MUMPOWER
PHILLIP NEWALL
TAKAO NOGUCHI
LUANA NUNES
KARSTEN OLSEN
PHILLIP PÄRNAMETS
PAULA PARPART
SALLYANN PARRY
ALVIN PASTORE
GILES PAVEY
ELIZABETH PAUL
REBECCA PIKE
ANGELO PIRRONE
GERIT PFUHL
GANNA POGREBNA
LIAM POLLOCK
ANDREA POLONIOLI
DIANE POMEROY
ROSIE PRIOR
MIKAEL PUURTINEN
TIM RAKOW
ANTONIO RANGEL
PETER RIEFER
JOHANNES RITTER
DAVID RONAYNE
ADAM SANBORN
JEFF SCHALL
MICHELLE SCHALL
ANYA SKATOVA
DYLAN SOUBIEN
SARAH SMITH
TOM STAFFORD
SANDRA STARKE
JACK STECHER
HRVOJE STOJIC
NEIL STEWART
BRIONY SWIRETHOMPSON
PETROC SUMNER
FLORIAN TANDONNET
PETE TRIMMER
JAMES TRIPP
KONSTANTINOS TSETSOS
HENK VAN DEN BERG
RONALD VAN DEN BERG
ERICJAN WAGENMAKERS
LUKASZ WALASEK
LEONARDO WEISSCOHEN
JAN K. WOIKE
DANIEL WOLPERT
JIAXIANG ZHANG
DMB 2014 | 912 Sept 2014 | 5
6 Information
8 Conference Schedule
13 Gerd GigerenzerSimple Heuristics for a Complex World
14 Contributed Talks: Day 1
33 EJ WagenmakersReinforcement Learning
and the Iowa Gambling Task
35 Jeff SchallDecisions, accumulators and neurons:
How secure a bridge?
36 Contributed Talks: Day 2
57 Nick ChaterVirtual bargaining: A microfoundation
for social decision making and interaction
59 Tom GriffithsPriors and processes
in Bayesian models of cognition
60 Contributed Talks: Day 3
81 Daniel WolpertProbabilistic models of sensorimotor
control and decision making
83 Antonio RangelThe neuroeconomics of simple choice
84 Contributed Talks: Day 4
98 Poster Session 1
107 Poster Session 2
6 | DMB 2014 | 912 Sept 2014
Conference VenueThe majority of DMB 2014 activities take place in
the Bristol science museum, AtBristol, located onAnchor Rd, Harbourside, Bristol BS1 5DB (see mapreference 1). The exception to this is the conferencedinner, which will take place in the Great Hall of thebeautiful Wills Memorial Building, Queen's Road (mapreference 2).
Practical Information relating to The Venue
Parking: The nearest car park is Millennium SquareCar Park, a safe and secure car park adjacent to AtBristol. It is an underground car park, open 24 hours aday with on site security and CCTV cameras. The carpark is situated underneath Millennium Square just a fewminutes walk from AtBristol. There is parking for 550cars. There is a 2m height restriction. Using Sat NavEnter BS1 5LL as your destination.
Toilets: As you enter the main Rosalind Franklin Roomthe toilets are through the double doors at the far endof the room. There are two accessible toilets, one ateach end of the room.
Fire Alarm: If the fire alarm is activated a loud sirenwill sound. Please leave via the pink or purple staircaseat each end of the room, following the green fire exitsigns. AtBristol staff will guide guests out of the buildingto the assembly point. The assembly point is on AnchorSquare (the cobbled square in front of the AtBristolbuilding). AtBristol staff will inform guests when it is safeto reenter the building.
Terraces: There is a smoking area on the NorthTerrace. Smoking is not allowed anywhere else withinAtBristol. Please note terraces become slippery whenwet. If in use please take care.
Registration Desk
The registration desk can be found in the entranceto the Rosalind Franklin Room, situated on the first floorof AtBristol. Please enter AtBristol through the main,ground floor entrance. You will be directed to thestairs or lift to the first floor. The registration desk willopen at 9am on Tuesday 9 September and will remainopen for late registrations and queries throughout theconference. A cloakroom facility will also be available.
Conference Name BadgeParticipants are kindly asked to wear the
conference badge at all times during the conference.It entitles them to participate in all activities and isrequired for entry to the Great Hall for the conferencedinner.
WiFi Internet AccessFree wireless internet access is available in the
conference venue. You can connect by enteringWiFi Code: AtBristol EventsPassword: MaxhertZ!WE11If you have any problems connecting, please go to
the registration desk.
Coffee Breaks and LunchesCoffee breaks and lunches are served in the
Maurice Wilkins Room. The timings of lunches vary duringthe conference so please check the schedule carefully.
Conference DinnerThe conference dinner will be held on Tuesday, 9th
September in the beautiful Wills Memorial building (mapreference 2). Delegates are invited to arrive at 19h15for a welcome drink. Guests will be seated at 19h45;dinner will be served at 20h00.
If you are walking, the map shows how to get thereon foot from the conference venue. There is no parkingat the venue itself; limited parking can be found in thestreets surrounding the venue and at Berkeley Square.
There are several bus services which stop in the ParkStreet / Triangle / Whiteladies Road area including the1, 2, 8, 9, 40, 40A and 41. Leave the bus at theBerkeley pub on Queen's Road which is opposite WillsMemorial Building.
Drinks ReceptionThe drinks reception will be held on Thursday, 11th
September, directly following the last talk of the day.Delegates are asked to make their way down to theground floor where they will also be able to view theexhibits at AtBristol.
Instructions to Presenters
Talk Presentations
Talks last no longer than 22 minutes, followed by 8minutes of discussion. These time limits will be strictlyenforced so that participants can switch sessions tohear specific talks. Each session will have a dedicatedlaptop that speakers can use to connect to theprojector. Speakers will also be able to connect theirown machines to the projector using a VGA cable. Iftheir laptops require any other connector (HDMI,Display port, etc), the speakers are required to bringtheir own adapter to a VGA connector. Please bringyour presentation on a USB stick a minimum of 10minutes before the beginning of the session. If usingyour own laptop, please bring it and test yourpresentation 15 minutes before the session starts.
DMB 2014 | 912 Sept 2014 | 7
Poster Presentations
Poster sessions will be held on Wednesday andThursday 15h30 – 16h45. Poster boards will be set upin the Rosalind Franklin Room. Each poster board will benumbered and you are asked to put your poster on theallocated board. Material to attach your poster will beprovided. On the day of your presentation please putup your poster by 10h30 and remove it after the lastsession of the day.
Practical Information
How to get to the conference
By Air: The nearest airport to the venue is BristolInternational Airport. From the airport, you can eithertake the bus (Bristol Flyer Express LInk) to Bristol CoachStation or a taxi (Checker Cars) directly to AtBristol. Ifyou're flying to London Heathrow, you can either get aNational Express bus directly to Bristol Coach Stationor train (via Paddington) to Bristol Temple Meads.
By Train: The city of Bristol is served by two majortrain stations Bristol Temple Meads and BristolParkway. The nearest station to the venue is BristolTemple Meads, Bristol BS1 6QF. From the train stationyou can either get a taxi to AtBristol or any of the FirstBuses that come to the City Centre and walk from theCentre to AtBristol (23 minutes on foot).
By Bus: The closest bus station is Bristol CoachStation on Marlborough Street (top right corner of mapon the inside front cover). You can walk from the coachstation to AtBristol (1015 minutes on foot) or take ataxi available at the exit of the coach station.
Getting Around
Walking: We find the best way to get around town isto walk. AtBristol is at the centre of town and a walkingdistance from public transport, cafés and hotels.
Bus: Several bus companies run buses around Bristol.On all of them, you pay the driver when you get on.Route details and timetable can be found usinghttp://www.travelwest.info/bus.
Taxi: Blue taxi cabs can be taken without prebooking. These can be found at various points aroundthe city (including the train station).
Private hire: VCars: 0117 925 2626
Emergency
Emergency Phone Number: 999. This number goesthrough to the police, the fire brigade and ambulanceservices. The nearest accident and emergencydepartment is situated at the Bristol Royal Infirmary,Marlborough Street, Bristol, Avon, BS2 8HW.
Non urgent medical advice can be sought from theNHS by calling 111
About Bristol
Attractions
We hope you will have time to explore the beatifulcity of Bristol whilst you are here. To find out whatBristol has to offer, take a look at http://visitbristol.co.uk.
Tourist information: Bristol Visitor Information Centre islocated on the waterfront at the E Shed, 1 CanonsRoad, Bristol, BS1 5TX (open: MonSat 10am – 5pmand Sun 11am – 4pm).
Eating Out
There are some amazing restaurants in Bristol wehave awardwinning restaurants, restaurants on boats,gastro pubs, cafés and bistros of all varieties. You willfind a wealth of eateries within walking distance of thevenue in Millennium Square and along the water frontas well as on Park street and Whiteladies Road.
Our Favourite Restaurants
Bordeaux Quay: VShed, Cannons Way, BS1 5UH
Cherry Duck: 3 Queen Quay, Welshback, BS1 4SL
Glass Boat: Welsh Back, BS1 4SB
Hotel du Vin: Narrow Lewins Mead, BS1 2NU
Sergio's: 13 Frogmore Street, BS1 5NA
Spyglass: Welsh Back, BS1 4SB
The Olive Shed: Princes Wharf, BS1 4RN
The Stable (good pizza): Harbourside, BS1 5UH
Our Favourite Cafés
Arnolfini: 16 Narrow Quay, BS1 4QA
Boston Tea Party: 75 Park St, BS1 5PF
Falafel King (van): Harbourside, City Centre
Mud Dock: 40 The Grove, BS1 4RB
The Folk House: 40A Park St, BS1 5JG
Watershed: 1 Canons Road, BS1 5TX
Our Favourite Bars/Pubs
Grain Barge: Mardyke Wharf, Hotwell Rd, BS8 4RU
Harveys Cellars: 12 Denmark St, BS1 5DQ
Llandoger Trow: King St, BS1 4ER
Milk Thistle (book ahead): Colston Avenue, BS1 1EB
The Cornubia (live music): 142 Temple St, BS1 6EN
The Cottage Inn (by boat): Cumberland Rd, BS1 6XG
The Hope & Anchor: 38 Jacob's Wells Rd, BS8 1DR
The Old Duke (for Jazz): 45 King St, BS1 4ER
The Ostrich Inn: Lower Guinea Street, BS1 6TJ
The White Lion: Sion Hill, BS8 4LD
Live Music
Fleece: http://www.thefleece.co.uk/
Louisiana: http://www.thelouisiana.net/
8 | DMB 2014 | 912 Sept 2014
09h00 10h30 Registration Opens
10h30 10h45 Welcome Remarks
10h45 11h45
11h45 12h00
12h00 12h30
Adam SanbornDecision making with ordered
discrete responses
Rong GuoAmygdala activity correlates withthe transition from saliencybased
to rewardbased choices
John CookRational irrationality: modeling
climate change beliefpolarization using Bayesian
networks
12h30 13h00
Gilles DutilhThe adaptive shifts of decision
strategies
Nishantha MalalasekeraNeuronal correlates of value
based decision making differ withinformation gathering strategies
Joanna M. BurchBrownDiversification, risk and climatechange: Towards a principled
guide to portfolio optimisation inclimate investment
13h00 14h15 Lunch
14h15 14h45
Luana NunesEthological decision making with
nonstationary inputs usingMSPRTbased mechanisms
Leonardo WeissCohenIncorporating conflicting
descriptions into decisions fromexperience
Jan K. WoikePyrrhic victories in publicgoods
games: when relative comparisonsmatter more than absolute
payoffs
14h45 15h15
Gaurav MalhotraChanging decision criteria in
rapid & slow decisions:do people behave optimally?
Jack StecherDescription and ExperienceBased Decision Making: AnExperimental and StructuralEstimation Approach to theDescriptionExperience Gap
Andrea IsoniPreference and Belief Imprecision
in Games
15h15 15h45
Rani MoranThe Collapsing ConfidenceBoundary Model: A Unified
Theory of Decision, Confidence,and ResponseLatencies
Thomas HillsChoiceRich Environments InduceRiskTaking and Reduce Choice
Deferral
Liam PollockReduced uncertainty improveschildren’s coordination, but not
cooperation
15h45 16h15
Petroc SumnerCompetitive accumulator models
explain why there is often nocorrelation between error and RT
measures of inhibitory control
Janine Christin HoffartThe Influence of Sample Size on
Decisions from Experience
Philip NewallFailing to Invest by the Numbers
16h15 16h45 Coffee Break
16h45 17h30 New Perspectives: EricJan Wagenmakers
19h15 23h00 Conference Dinner, Wills Memorial Building
Chair: John McNamara
Chair: Rafal Bogacz Chair: Casimir Ludwig Chair: Joanna BurchBrown
Keynote: Gerd Gigerenzer
Coffee Break
Chair: Rafal Bogacz Chair: Andreas Jarvstad Chair: David Leslie
Rosa l i nd Frank l i n Room Franc i s Cr ick Room James Watson Room
DMB 2014 | 912 Sept 2014 | 9
09h00 10h00
10h00 10h30
10h30 11h00
Eugene McSorleyDissociation between the impactof evidence on eye movementtarget choice and confidence
judgements
Neil BramleyActing informatively How people
learn causal structure frominterventions
Florian ArtingerRisk preferences in context
11h00 11h30
Andreas JarvstadThe neural substrate of eyemovement control: Inhibition,
selection and choice
Hrvoje StojicFunction learning while making
decisions
Andrea PolonioliStanovich's challenge to the
"adaptive rationality" project: anassessment
12h30 14h00 Lunch
14h00 14h30
Jiaxiang ZhangDecisionmaking deficits in
neurodegenerative diseases
Gordon D. A. BrownRelative Rank Theory
Ullrich EckerThe Effects of Misinformation onReasoning and Decision Making
14h30 15h00
Encarni MarcosThe variance of neuronal activityin dorsal premotor area predicts
behavioral performance: Themodulation of behavior andneural activity by trial history
Tatiana KornienkoNature's Measuring Tape:
A Cognitive Basis forAdaptive Utility
Sudeep BhatiaNaturalistic MultiAttribute Choice
15h00 15h30
Gilles de HollanderLarge Cortical Networks in a
Small Nucleus: a 7T fMRI Study onLimbic, Cognitive and Associative
Subparts in The SubthalamicNucleus During Decisionmaking
Sarah SmithIs financial risk attitude entirely
relative?
Julia BrettschneiderDecisionmaking in adjuvant
cancer treatment choices basedon complex information, including
genomic recurrence risk scores
15h30 16h45 Poster Session 1 (Coffee Available)
16h45 17h30 New Perspectives: Nick Chater
Chair: Iain Gilchrist
Chair: Iain Gilchrist Chair: Jeff Bowers Chair: David Leslie
Keynote: Jeff Schall
Coffee Break
Chair: Iain Gilchrist Chair: Roland Baddeley Chair: Steve Lewandowsky
Rosa l i nd Frank l i n Room Franc i s Cr ick Room James Watson Room
11h30 12h00
Aline BompasA unified framework for speededsaccadic and manual decisions
Alice MasonChance based uncertainty of
reward improves longtermmemory
Ganna PogrebnaAmbiguity attitude in
coordination problems
12h00 12h30
Philip PärnametsEye gaze reflects and causes
moral choice
Tim Fawcett'Irrational' behaviour can arisefrom decision rules adapted to
complex environments
David KelseyAmbiguity in Coordination Games
10 | DMB 2014 | 912 Sept 2014
09h00 10h00
10h00 10h30
10h30 11h00
Bradley C. LoveOptimal Teaching with Limited
Capacity Decision Makers
Shenghua LuanThresholdbased Inference
Christophe TandonnetDecision making and
response implementation:the missing link to constrain
models
11h00 11h30
Adnane EzziziLearning to solve
working memory tasks
Konstantinos TsetsosThe value of economic
irrationality: why and when beingirrational is advantageous
Valerio BiscioneRate Domain Distributions inSimple Reaction Time Task
12h30 14h00 Lunch
14h00 14h30
Timothy MullettVisual attention in choice: Do
individuals pay more attention tomore important information?
Elliot LudvigBig, fast, and memorable:
Increased gambling in riskydecisions from experience
Claire HalesValidation of the diffusion model
for investigating the decisionmaking processes underlyingcognitive affective bias in
rodents
14h30 15h00
Rosie ClarkThe effect of spatial probability
and reward on response time
Lukasz WalasekLoss aversion is a property of the
experimental design, not theparticipant
Pete TrimmerOptimal moods in behavioural
models
15h00 15h30
Luigi AcerbiA matter of uncertainty: Optimalityand suboptimality in sensorimotor
learning
David RonayneMultiAttribute Decisionby
Sampling
Carolina DoranEconomical decision making by
Temnothorax albipennis antcolonies
15h30 16h45 Poster Session 2 (Coffee Available)
16h45 17h30 New Perspectives: Daniel Wolpert
Chair: David Leslie
Chair: Kevin Lloyd Chair: Tim Fawcett Chair: Gaurav Malhotra
Keynote: Tom Griffiths
Coffee Break
Chair: Casimir Ludwig Chair: Gaurav Malhotra Chair: Andrew Higginson
Rosa l i nd Frank l i n Room Franc i s Cr ick Room James Watson Room
11h30 12h00
Carolina Feher da SilvaA Simple Method to Characterize
the ExplorationExploitationTradeOff in Binary Decision
Making
Perke JacobsA Competitive Test of SatisficingHeuristics in Choice From Serially
Dependent Sequences
Angelo PirroneEvolutionary and computational
considerations on the DriftDiffusion Model
12h00 12h30
John GroganDopamine affects encoding andretrieval of positive and negative
reinforcement learning
Paula ParpartHeuristics as a special case of
Bayesian inference
Takao NoguchiThe Dynamics of EvidenceAccumulation and Choice
17h30 19h30 Drinks reception and @Bristol exhibits
DMB 2014 | 912 Sept 2014 | 11
09h00 10h00
10h00 10h15
10h15 10h45
Daniel A. BraunEllsberg's paradox
in sensorimotor learning
Russell GolmanPolya's Bees: A Model of
Decentralized Decision Making
10h45 11h15
A. Aldo FaisalA learning rule for optimal
feedback control that predictscontinuous action selection in
complex motor tasks
Peter S. RieferCoordination in large groups:What would we do if zombies
attack?
11h15 11h30 Coffee Break
11h30 12h00
Rafal BogaczTo act or not to act: Learning the
value of not acting
Andrew HowesComputationally Rational
Decision Making
Dan BangConfidence matching in social
interaction
12h00 12h30
M.C. KeukenThe role of the subthalamic
nucleus in multiple alternativeperceptual decisionmaking
revealed by 7T structural andfunctional MRI
Laurence HuntSuboptimal information search
strategies duringmultiattribute choice
Rebecca FloydBetter Together UnderstandingCollaborative Decision Making
12h30 13h00
Nathan LeporaPerceptual decision making in
robotics and neuroscience
George FarmerThe attraction effect is rationalgiven uncertainty in expected
value calculation
Karsten OlsenKeeping track of others' choices:
Predicting changes in theperceptual decisions and
decision confidenceof other individuals
13h00 13h15 Closing remarks
Chair: Casimir Ludwig
Chair: Roland Baddeley Chair: Pete Trimmer
Keynote: Antonio Rangel
Coffee Break
Chair: Nathan Lepora Chair: Joanna BurchBrown Chair: John McNamara
Rosa l i nd Frank l i n Room Franc i s Cr ick Room James Watson Room
12 | DMB 2014 | 912 Sept 2014
DMB 2014 | 912 Sept 2014 | 13
Rosa l i nd Frank l i n Room
GERD GIGERENZER Max Planck Institute for Human Development;Director of the Harding Center for Risk Literacy in Berlin
How to invest? Whom to trust? Complex problems requirecomplex solutions – so we might think. And if the solution doesn ’twork, we make it more complex. That recipe is per fect for a worldof known risks, but not for an uncertain world, as the failure of thecomplex forecasting methods leading to the 2008 financial crisisil lustrates. In order to reduce estimation error, good inferencesunder uncertainty counterintuitively require ignoring part of theavailable information. Less can be more. Yet although we facehigh degrees of uncertainty on a daily basis, most of economicsand cognitive science deals exclusively with lotteries and similarsituations in which all risks are per fectly known or can be easilyestimated. In this talk, I invite you to explore the land ofuncertainty, where mathematical probability is of limited value andpeople rely instead on simple heuristics, that is, on rules of thumb.We meet Homo heuristicus, who has been disparaged by manypsychologists as irrational for ignoring information—unlike the morediligent Homo economicus. In an uncertain world, however, simpleheuristics can be a smart tool and lead to even better decisionsthan with what are considered rational strategies. The study ofheuristics has three goals. The first is descriptive: to analyze theheuristics in the “adaptive toolbox ” of an individual or aninstitution. The second goal is normative: to identify theecological rationality of a given heuristic, that is, the structures ofenvironments in which it succeeds and fails. The third goal isengineering: to design intuitive heuristics such as fastandfrugaltrees that help physicians make better decisions.
Day 1 (Tuesday) 10h45
GERDGIGERENZER
14 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
ADAM SANBORN University of WarwickULRIK BEIERHOLM University of Birmingham
Analyses of decisionmaking behavior have often compared peopleagainst a normative standard, assuming that people are attempting tomaximize their expected rewards. The Bayesian formulation of this problemprescribes what should be done: people’s prior beliefs are combined withthe likelihood of the observed stimulus given the response via Bayes rule toproduce posterior beliefs about which response is correct. The posteriorbeliefs are then integrated with the loss function, which describes the gain invalue for a given choice for each possible true answer, to determine whichresponse has the highest expected value.
When people are asked to choose from among a small set of discreteresponses, the most straightforward approach is to treat the responses asunrelated, allowing the use of a multinomial prior distribution. The lack oforder to the responses calls for an allornone loss function that prescribespicking the response with the highest posterior probability. However, whenasked to make a continuous response, order is very important. People’sresponses are often modeled with normal distributions for the prior andlikelihood, which leads to a normal posterior. Most loss functions prescribechoosing the mean of the posterior.
Lying between these two wellexplored cases is decision making withordered discrete responses. This kind of decisionmaking is prevalentoutside the laboratory. People are often asked to make discrete responsesabout variables that are experienced as continuous: On what day thisweek will you finish the project? Ordered discrete responses particularlyarise when it comes to counting numbers of objects. Here the ground truth,such as the number of horses in a paddock, is irreducibly discrete and adiscrete response is necessary.
We investigated how people make decisions with a small number ofordered discrete responses, using a numerosity task in which participantsare asked to count the number of dots that were shown in a very brieflypresented display. We characterized the kinds of prior distributions,likelihood distributions, and loss functions that people used in this task.People’s choices were not well described by either common discrete orcommon continuous models of normative decisionmaking. The likelihoodsand loss functions reflected the ordered nature of the responses, butpeople learned prior distributions that reflected the discrete nature of thetask. Hence the best explanation of people’s decision making with ordereddiscrete responses involved aspects of both approaches.
Keywords: Bayesian decision making, ordered discrete responses, numerosity
Day 1 (Tuesday) 12h00
DMB 2014 | 912 Sept 2014 | 15
Franc i s Cr ick Room
RONG GUO Technische Universitat, BerlinKLAUS OBERMAYER University Medical Center, HamburgEppendorfJAN GLÄSCHER Bernstein Center for Computational Neuroscience, Berlin
Introduction: Ten years ago a finding that the ventral striatum (vStr)responds to salient, but nonrewarding stimuli (Zink et al. (2003))challenged the hitherto accepted theory of the vStr as one of the criticalhubs in the decisionmaking network encoding rewards and rewardprediction errors (O'Doherty (2004)). Although the coexistence bothreward and saliencyrelated signals in vStr has ever since been confirmedthe precise interaction of the two signals has not been fully resolved. Here,we approach this question computationally using reinforcement learningtheory in the context of modelbased fMRI in a task that requires subject tomake probabilistic choices for salient visual stimuli that may or may not leadto a subsequent reward.
Methods: 27 participants were instructed to predict the occurrence of avisual stimulus by choosing between left and right button presses in thescanner. Stimuli appeared probabilistically on the left and right screen sidein two conditions: (a) equal stimulus probability of 0.5 or (b) biased stimulusprobabilities of 0.7 and 0.3 counterbalanced between runs. Conditionalreward occurred probabilistically with 0.2 and 0.8 for correct predictionson the left and on the right side. Critically, in the biased stimulus condition,higher reward probability was assigned to the lower stimulus probability,thus creating a situation of disinformation, in which the salient visual stimuluswas not predictive of higher reward probabilities. We modeled choicebehavior with two independent reinforcement learning algorithms (RescorlaWagner models for stimulus and reward occurrences) that were negotiatedby a decaying tradeoff parameter η. We extracted modelderived signalsfor η, stimulus and reward prediction error. These signals are used in amodelbased fMRI analysis.
Results: Modelfree behavioral analyses confirmed that the (misleading)side with higher stimulus probability in the biased condition was chosenmore frequently than in the equal condition. This was also confirmed throughmodelbased analysis, in which the tradeoff parameter η decayed morequickly in the equal condition suggesting a faster transition to purelyrewardbased choices than in the stimulus condition. At the neural level wefound a coexistence of stimulus and reward prediction errors in the ventralstriatum suggesting that this region responses to surprising perceptualevents as well as unexpected reward delivery or omission. Furthermore, theamygdala correlated with the tradeoff parameter η.
Conclusions: Our findings suggest that vStr is responsive to expectancyviolations both at the pure perceptual and at the reward level. This isconsistent with previous studies that posit a dual role for vStr in rewardlearning and saliency detection. Furthermore, the amygdala appears to benegotiating between an initial emphasis on choosing the salient stimulusand pure rewardbased choices later. This may reflect its prominent role inPavlovian stimulusoutcome association, which in our study have to beovercome, in order to maximize the payoff. In summary, our studies highlightsthe roles of key parts of the decisionmaking network in learning stimulusand rewardbased choices and trading off both associations against eachother.
Keywords: Reinforcement learning, modelbased fMRI, reward
Day 1 (Tuesday) 12h00
16 | DMB 2014 | 912 Sept 2014
James Watson Room
JOHN COOK University of Queensland; University of Western AustraliaSTEPHAN LEWANDOWSKY University of Bristol; University of Western Australia
Belief polarization is said to occur when two people respond to thesame evidence by updating their beliefs in opposite directions. Thisresponse is considered irrational because it appears to violate normativelyoptimal belief updating, as for example dictated by Bayes ’ theorem. In lightof much evidence that people are capable of normativelyoptimal decisionmaking, belief polarization seemingly presents a puzzling exception. Weshow using Bayesian Networks that belief polarization can occur even ifpeople are making decisions rationally. We develop a computationalcognitive model that uses Bayes Nets to simulate Bayesian beliefpolarization. The Bayes Net model is fit to experimental data involvingrepresentative samples of Australian and US participants. AmongAustralians, intervention text about the scientific consensus on climatechange partially neutralised the influence of ideology, with conservativesshowing a greater increase in acceptance of humancaused globalwarming relative to liberals. In contrast, consensus information had apolarizing effect among the U.S. participants, with a reduction inacceptance of humancaused global warming among conservatives. FittingBayes Net models to the survey data indicates that a heightenedexpectation of fraudulent behaviour by climate scientists drives the backfireeffect among U.S. conservatives.
Keywords: Belief polarization, Bayesian Networks, climate change, scientificconsensus
Day 1 (Tuesday) 12h00
DMB 2014 | 912 Sept 2014 | 17
Day 1 (Tuesday) 12h30Rosa l i nd Frank l i n Room
GILLES DUTILH University of BaselBENJAMIN SCHEIBEHENNE University of BaselHAN L. J. VAN DER MAAS University of AmsterdamJÖRG RIESKAMP University of Basel
In many domains of cognitive psychology, it is proposed that peopleare equipped with a repertoire of strategies to solve the problems theyface. For instance, it is proposed that when people have to make aprobabilistic inference based on the information of multiple cues, theysometimes rely on simple heuristics (like the Take The Best rule) andsometimes apply strategies that require more elaborate processing. Indeed,many studies suggest that people's inferences can be described bydifferent strategies in different situations. However, critics of this view suggestthat people do not apply different strategies in each situation but insteadadjust one single strategy to the characteristics of each situation. In thisstudy we examine the strategies that individuals use when availableresources change. Specifically, we continuously changed the relativeimportance of speed and accuracy in a multiplecue inference task. Bydoing so, participants where incentivized to gradually shift from very fastguessing behavior to slow but highly accurate behavior (and back andforth). Thorough investigation of participants' behavior at the transitionsbetween fast and accurate behavior offers insight into whether peopleselect different strategies or rather continuously adjust one single strategyto meet varying task requirements. Results show that participants are ableto adjust behavior to some extent, but also display qualitative shifts inbehavior.
Keywords: MultiAttribute Decision Making, Strategy Selection, SpeedAccuracyTradeOff.
18 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 12h30Franc i s Cr ick Room
NISHANTHA MALALASEKERA UCL Institute of Neurology, LondonLAURENCE T HUNT UCL Institute of Neurology, LondonBRUNO MIRANDA UCL Institute of Neurology, LondonTIMOTHY E BEHRENS UCL Institute of Neurology, LondonSTEVE W KENNERLEY UCL Institute of Neurology, London
Optimal decisionmaking depends on gathering sufficient information todetermine the best outcome. However, multiattribute decisions present theadditional challenge of deciding what information to use to compare thealternatives. Patients with damage to the ventromedial prefrontal cortex(vmPFC) not only exhibit deficits in valuebased decisionmaking, they alsouse different information to make choices; while healthy humans tend toanalyse choices using a withinattribute comparison strategy (e.g.,comparing the price of each house), patients with vmPFC damage tend touse a withinoption strategy (e.g., determine the size, cost, etc. of individualhouses). These findings suggest a critical function of vmPFC may be toregulate the process by which choices are compared. Other animal andhuman lesion studies have described a double dissociation betweenanterior cingulate cortex (ACC) and orbitofrontal cortex (OFC) damagewhere ACC lesions cause specific deficits in action based decisionmakingwhereas OFC lesions profoundly affect stimulus based decisions. Theseresults suggest that different parts of prefrontal cortex may representrelevant decision variables in different frames of reference.
To study multiattribute value based decisionmaking, monkeys were firsttaught a set of picture value associations for each of two attributes(reward probability, reward size; 5 pictures per attribute). Subjects were thenpresented with two pictures (one from each attribute) on each side of afixation point (4 pictures in total) and made choices between left and rightoptions. Importantly, subjects were free to saccade to the different picturesto gather information about attributes/options. Subjects were also free toindicate their choice (by moving a joystick to the left/right) withoutnecessarily viewing all the picture information. Eye movements thereforeprovided a proxy for the information gathering strategies influencingdecisionmaking. Each cue consisted of two important properties other thanits value; firstly its position on the screen (left or right) which was tied to theaction required to obtain the option. Secondly, the cue can denote eithera probability or magnitude attribute.
Single neurons were recorded from ACC, OFC, dorsolateral PFC(DLPFC) and vmPFC while subjects performed the task. When subjects werepresented with the first picture cue in the trial, many of neurons throughoutall four brain areas encoded its value. This encoding was substantiallystronger and earlier in ACC and OFC than vmPFC and DLPFC. However, asignificant subpopulation of ACC and DLPFC neurons differentiallyencoded the value of the cue when presented on the left compared to theright suggesting an action value frame of reference for these neurons. Incontrast, a significant subpopulation of OFC neurons differentially encodedthe value of the cue depending on whether it was of probability ormagnitude type. Analysis of neuronal responses to subsequent cues havefound that OFC and ACC maintain this dissociation of value referenceframes throughout the trial but value signals evolve from encoding the valueof what is currently presented to encoding the value of what will beeventually chosen.
Keywords: Decision Making, Information Gathering, Prefrontal Cortex, AnteriorCingulate Cortex, Orbitofrontal Cortex, Ventromedial Prefrontal Cortex
DMB 2014 | 912 Sept 2014 | 19
Day 1 (Tuesday) 12h30James Watson Room
JOANNA M. BURCHBROWN University of Bristol
Climate policies are likely to have important sideeffects that policymakers and investors will fail to anticipate in advance. For example, theWorld Bank estimates that rising food prices in 2010 drove an additional44 million people below the poverty line in Africa, and research suggeststhat increased demand for biofuels was an important factor contributing tothe rise in food poverty. Demand for biofuels can also lead to destructionof important habitats. In Indonesia, over nine million acres of rainforesthave been cleared in recent decades to support palm oil plantations.Most of this production has historically been used for food, but palm oil isalso one of the main biofuels, and projections suggest that increaseddemand for biofuel in the next decade is likely to lead to further clearing oflarge areas of rainforest in South East Asia and elsewhere. Althoughbiofuels are often advocated as an important technology for emissionsreduction, they have become more controversial in recent years as a resultof these adverse sideeffects.
The possibility that climate investment may have adverse side effects isimportant for two reasons. One is that adverse sideeffects risk underminingthe goals that justify the policies in the first place (eradication of poverty,social and economic development, protection of biodiversity and habitat,etc). The second is that perceptions of adverse sideeffects can lead towithdrawal of political will for climate mitigation. This in turn increases 'policyrisk' instability in policy support for green investment. Potential investorstypically identify policy risk as one of the major blocks to current investmentin green technologies.
What strategies can policymakers and investors adopt to mitigate therisk of adverse sideeffects, given that they have good reason to believethat they must make their decisions without full understanding of thesystematic effects of their actions?
In this paper, I argue that we can substantially reduce risk of adverseconsequences from climate investment by drawing on key principles of riskmanagement from the financial theory of portfolio analysis. I begin byexplaining three central principles of portfolio theory. I then show how wemight apply these principles to manage risk in relation to climate policy andinvestment. I argue in particular that we should systematically usediversification to reduce risk in our responses to climate change. I explainsome of the challenges for figuring out whether we have successfullydiversified risk in this context; and suggest some simple heuristics that wemight apply to make the theory useful in practice for climate policy makersand investors.
Keywords: Climate change, risk, uncertainty, portfolio theory, risk management,climate ethics
20 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
NISHANTHA MALALASEKERA UCL Institute of Neurology, LondonLAURENCE T HUNT UCL Institute of Neurology, LondonBRUNO MIRANDA UCL Institute of Neurology, LondonTIMOTHY E BEHRENS UCL Institute of Neurology, LondonSTEVE W KENNERLEY UCL Institute of Neurology, London
Optimal decisionmaking depends on gathering sufficient information todetermine the best outcome. However, multiattribute decisions present theadditional challenge of deciding what information to use to compare thealternatives. Patients with damage to the ventromedial prefrontal cortex(vmPFC) not only exhibit deficits in valuebased decisionmaking, they alsouse different information to make choices; while healthy humans tend toanalyse choices using a withinattribute comparison strategy (e.g.,comparing the price of each house), patients with vmPFC damage tend touse a withinoption strategy (e.g., determine the size, cost, etc. of individualhouses). These findings suggest a critical function of vmPFC may be toregulate the process by which choices are compared. Other animal andhuman lesion studies have described a double dissociation betweenanterior cingulate cortex (ACC) and orbitofrontal cortex (OFC) damagewhere ACC lesions cause specific deficits in action based decisionmakingwhereas OFC lesions profoundly affect stimulus based decisions. Theseresults suggest that different parts of prefrontal cortex may representrelevant decision variables in different frames of reference.
To study multiattribute value based decisionmaking, monkeys were firsttaught a set of picture value associations for each of two attributes(reward probability, reward size; 5 pictures per attribute). Subjects were thenpresented with two pictures (one from each attribute) on each side of afixation point (4 pictures in total) and made choices between left and rightoptions. Importantly, subjects were free to saccade to the different picturesto gather information about attributes/options. Subjects were also free toindicate their choice (by moving a joystick to the left/right) withoutnecessarily viewing all the picture information. Eye movements thereforeprovided a proxy for the information gathering strategies influencingdecisionmaking. Each cue consisted of two important properties other thanits value; firstly its position on the screen (left or right) which was tied to theaction required to obtain the option. Secondly, the cue can denote eithera probability or magnitude attribute.
Single neurons were recorded from ACC, OFC, dorsolateral PFC(DLPFC) and vmPFC while subjects performed the task. When subjects werepresented with the first picture cue in the trial, many of neurons throughoutall four brain areas encoded its value. This encoding was substantiallystronger and earlier in ACC and OFC than vmPFC and DLPFC. However, asignificant subpopulation of ACC and DLPFC neurons differentiallyencoded the value of the cue when presented on the left compared to theright suggesting an action value frame of reference for these neurons. Incontrast, a significant subpopulation of OFC neurons differentially encodedthe value of the cue depending on whether it was of probability ormagnitude type. Analysis of neuronal responses to subsequent cues havefound that OFC and ACC maintain this dissociation of value referenceframes throughout the trial but value signals evolve from encoding the valueof what is currently presented to encoding the value of what will beeventually chosen.
Keywords: Decision Making, Information Gathering, Prefrontal Cortex, AnteriorCingulate Cortex, Orbitofrontal Cortex, Ventromedial Prefrontal Cortex
Day 1 (Tuesday) 14h15Rosa l i nd Frank l i n Room
LUANA NUNES University of SheffieldKEVIN GURNEY University of Sheffield
The Multihypothesis Sequential Probability Ratio Test (MSPRT) is anasymptotically optimal, sequential test for decision making between severalalternatives. It works by integrating ‘evidence’ on each of its ‘channels ’ andmakes a decision when the first channel output crosses a threshold. Theoverwhelming majority of its applications focus on making single, discretedecisions with stochastically stationary inputs supplying evidence, and inwhich the evidence is initialized to zero prior to the decision process. Here,the decision making process is terminated upon crossing the threshold,which therefore constitutes an absorbing boundary. In contrast,ethologically plausible decision making in animals (and situated agents)should proceed in a continuous fashion, and will occur in environmentswhere the information supporting each alternative changes dynamicallythroughout the integration process; that is the inputs will be nonstationary.Models of decision making and action selection in animals are often basedon the hypothesis that the basal ganglia – a set of subcortical brain nuclei– act as the central mechanism mediating decisions. These models work withcontinuously varying inputs, and select actions (make decisions) via athreshold crossing which does not terminate the decision making process.Intriguingly, (Bogacz & Gurney, 2007) showed there was closecorrespondence between the functional anatomy of the basal ganglia andthe algebraic manipulations of the MSPRT, but considered only discretedecision trials in reporting their results. However, viewed as an algorithmicabstraction of an anatomical architecture, we ask the question: can theMSPRT mechanism be recruited for decision making with nonstationaryinputs, and without use of an adhoc ‘ reset ’ mechanism? We present apreliminary series of simulation experiments exploring this possibility, andcomparing the results with the full basal ganglia model (Gurney, Prescott, &Redgrave, 2001) working in continuous selection mode. In order to do this,we introduced what we call a transparent boundary, which records choicesmade but does not stop the integration. Further, as with the basal gangliamodel, threshold crossing from above comprises a selection, whereascrossing from below indicates its deselection. Using inputs in which two outof three channels had very similar means, the MSPRT mechanism showedbehavior consistent with that obtained from the basal ganglia model(Gurney, Prescott, & Redgrave, 2001); thus, closely matched competitorscan induce a ‘selection limiting ’ effect on each other and take more time toget selected than a channel with the same input but no close competitors.Continuous decision making of the kind we envisage here introduces aproblem: continuous evidence accumulation, results in unfeasibly largeinputs, but also in the inability to ‘ forget ’ past evidence and focus on morerecent input. We therefore, also introduced an integration window so thatonly the N most recent samples of information are used in the decisionprocess. With this device, timetoselect after introduction of a new inputsignal contrast, was considerably decreased because the ‘’baggage’ ofthe previous evidence history could be discarded more quickly. Takentogether, our results indicate how the MSPRT mechanism may be recruitedfor continuous, ethologically plausible decision making with nonstationaryinputs.
Keywords: decision making, multihypothesis, MSPRT, sequential test, continuousalgorithm, nonstationary inputs
DMB 2014 | 912 Sept 2014 | 21
Day 1 (Tuesday) 14h15Franc i s Cr ick Room
LEONARDO WEISSCOHEN University College LondonEMMANOUIL KONSTANTINIDIS University College LondonMAARTEN SPEEKENBRINK University College LondonNIGEL HARVEY University College London
Most of the research on the "descriptionexperiencegap'' so far hasfocused on presenting different participants with either descriptiononly orexperienceonly tasks separately. However, decisions in everyday life arealso commonly made on a combination of descriptive and experientialinformation, and these two sources of information often contradict eachother. Our aim was to shed additional light on how individuals combineinformation acquired by description and by experience when makingdecisions. Our experiments have shown that individuals exposed to bothdescription and experience can be influenced by the two sources ofinformation, in particular in situations in which the description was in conflictwith the experience. We looked at how individuals might integrateexperience with prior beliefs about the outcomes of their choices, withdifferent weights given to each source of information, depending on theirrelevance. Cognitive models that included the descriptive information fittedthe human data more accurately than models that did not includedescriptions. We discuss the wider implications for understanding how thesetwo commonly available sources of information can be combined for dailydecisionmaking, and the impact on areas such as warning labels.
Keywords: description, decisions from experience, descriptionexperience gap,reinforcement learning, repeated decisions
22 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 14h15James Watson Room
SEBASTIAN HAFENBRÄDL University of LausanneJAN K. WOIKE Max Planck Institute for Human Development
Participants in repeated publicgoods games do not fully contribute tothe public good, surprisingly even when the typical dilemma is absent andcontributions maximize both individual and group outcome. Here, wedemonstrate that destructive competition based on relative comparisons isthe driving factor for suboptimal (and spiteful) decisions in games where fullcooperation is optimal for the individual player. Adding social informationto the individual outcome feedback can destroy future selfbeneficialcooperation when focused on relative gains and rankings and help tomaintain cooperation levels in true dilemmas when focused on absolutegroup performance and efficiency. Ranking information triggered relativecomparisons that led to an erosion of cooperation, whereas spitefulbehavior is severely limited in the absence of relative information.Explanations centering on mere confusion, unconditional social preferencesor a universal cooperative motivation cannot account for these results. Ourfindings demonstrate that feedbackstructures can influence cooperationlevels and trigger social comparisons that lead to valuedestroyingcompetition. Designers of information environments inside and outside thelab need to be aware of these consequences. Furthermore,interpretations of behavior shown in economic games need to reflect thedemonstrated dependency of cooperative actions on highly variablefeatures of social environments.
Keywords: social comparisons, cooperation, economic games, confusion
DMB 2014 | 912 Sept 2014 | 23
Day 1 (Tuesday) 14h45Rosa l i nd Frank l i n Room
GAURAV MALHOTRA University of BristolDAVID LESLIE University of BristolCASIMIR LUDWIG University of BristolRAFAL BOGACZ University of Bristol; University of Oxford
Research on decision making has indicated that the tradeoff betweenthe speed and accuracy with which people make decisions is determinedby a decision criterion; a threshold on the evidence that needs to begathered before the decisionmaker commits to a choice. A fundamentalquestion that has attracted a lot of interest is whether humans decreasethe decision threshold within the course of a single decision. Weinvestigated this issue from both a normative and an empirical perspective.Using a simple theoretical framework, we determined the minimal conditionsunder which a decisionmaker should vary their decision criterion tomaximise reward rate. We found that the shape of the optimal thresholddepends crucially on how noise in signals varies across trials. Optimalthresholds stay constant within a trial when noise in sensory input isconsistent across trials. In contrast, when noise across trials is variable, timespent within a trial becomes informative and the optimal decision criterionmay not only decrease, but also increase or remain constant, dependingon the mixture of noise levels across trials.
We compared the performance of this optimal decisionmaker tobehaviour using a series of experiments. An expanded judgement paradigmthat precisely matched the theoretical framework showed that, in agreementwith the optimal model, participants had on average constant thresholds inblocks of trials with constant difficulty, and decreasing thresholds in blocksof trials with mixed difficulties. We then tested whether results from thisexpanded judgement paradigm generalise to faster decisions, likeperceptual decisions, with mean RTs below 1000ms. We found that theshape of decision boundaries critically depends on the specific mixture ofdifficulties only when participants believe that some trials in a block areextremely difficult, do they decrease their decision boundaries significantlyover the course of a decision.
Keywords: decreasing bounds, optimal decisions, reward rate, driftdiffusion model
24 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 14h45Franc i s Cr ick Room
JACK STECHER Carnegie Mellon UniversityLINDA MOYA Carnegie Mellon University
In experiments on decision making under risk, participants who learnprobabilities by sampling behave differently from those to whom theexperimenter describes probability distributions. This difference is known asthe descriptionexperience gap. We argue the chief reason for thedescriptionexperience gap is that participants who learn by samplingbecome familiar with a decision problem, prior to making a choice.Participants who learn from description do not have this luxury. We runlaboratory experiments, in which participants learn by sampling, and usetheir behavior to estimate the distribution of risk aversion coefficients. Adifferent group of participants receives equal task exposure, but samplesfrom a nonstationary, nonergodic distribution, after which they are told thedistributions in a choice problem. After controlling for task exposure, thegap disappears, and participants' behavior is shockingly consistent withexpected utility theory. We reanalyze older data, and find that our resultsare robust across contexts.
Keywords: decisions from experience; descriptionexperience gap; rare events; risk;probability weighting; utility; prospect theory
DMB 2014 | 912 Sept 2014 | 25
Day 1 (Tuesday) 14h45James Watson Room
DAVID BUTLER Murdoch School of Management and Governance, PerthANDREA ISONI Warwick Business SchoolGRAHAM LOOMES Warwick Business SchoolDANIEL NAVARROMARTINEZ Universitat Pompeu Fabra and BarcelonaGraduate School of Economics, Barcelona
Many experimental studies have found that the predictions of standardgame theory fail systematically when a game is faced by participants forthe first time (see Camerer 2003 for a review). Those predictions are basedon assuming that players are rational, selfinterested, risk neutral expectedutility maximisers, that rationality is common knowledge, and that the payoffsin the game represent utilities and are known precisely by all players. Themost standard solution concept, the Nash Equilibrium (NE), results in beliefsand strategy choices that are mutually consistent.
The failure of NE as a descriptive model of behaviour can be causedby failures of any of its underlying assumptions (or combination thereof). Anextensive branch of the gametheoretic literature has explored alternativesto the assumptions that players are selfinterested, in the form of socialpreferences of some kind (see Cooper and Kagel 2013 for a recentreview). Another branch has retained the general best response structurewhile relaxing the assumption that players have equilibrium beliefs in favourof some form of hierarchical beliefs, as in levelk or cognitive hierarchymodels (see Crawford et al. 2013 for a review). The ‘Quantal ResponseEquilibrium’ branch has retained the emphasis on equilibrium, while movingaway from the assumption that utilities are known with absolute precision(e.g. McKelvey and Palfrey 1995).
Motivated by the extensive evidence on the probabilistic nature ofchoice in individual decision making (e.g. Rieskamp et al. 2006), weinvestigate whether deviations from NE can be explained by empiricalmeasures of imprecision about players ’ strategy choices or about theirbeliefs.
We proxy preference imprecision using measures of confidence andelicit player ’s dispersion around their beliefs in a series of twelve 2x2games, which fall into two broad categories: games of conflict and gamesof cooperation. The games of conflict comprise three battleofthesexesand four matchingpennies games, whereas the games of cooperationcomprise three prisonersdilemma and two staghunt games. Within eachgame type, the payoffs in one of the cells are varied, while the others arekept constant, to detect systematic changes in beliefs and strategychoices.
We find substantial degrees of imprecision around both preferencesand beliefs. Our confidence measures are lowest in symmetric conflictgames with very unpredictable outcomes (e.g. a symmetric matchingpennies game), and highest in symmetric cooperation games with riskdominant equilibria (e.g. one of the staghunt games). Beliefs respondcoherently to parameter changes and vary in their degree of imprecisionshowing a systematic relationship with confidence. The rate at which playersbest respond to their beliefs is positively correlated with their confidence,but not with the dispersion of the beliefs themselves.
Keywords: preference imprecision, imprecise beliefs, experimental games.
26 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 15h15Rosa l i nd Frank l i n Room
RANI MORAN Tel Aviv UniversityANDREI R. TEODORESCU Tel Aviv UniversityMARIUS USHER Tel Aviv University
While recent research has made impressive progress in understandingthe mechanism underlying decisionmaking, decisionconfidence – animportant regulatory decision process – is still raising considerablechallenges. The choice followed by confidence paradigm produces richdata sets consisting of four important measures of cognitive performance:Choice, confidence and their corresponding latencies. These variablescombine and interact to form an extensive manifold of patterns thatchallenges confidence modelers. The resolution of confidence, the positivecorrelation between choiceaccuracy and choiceconfidence, is one suchpattern, centeral to our investigation. In two perceptual choice experiments,we manipulated the perceptual availability of the stimulus following thedecision and found systematic influences on the resolution of theconfidence judgments. This indicates that postchoice evidence exerts acausal effect on confidence and that the postchoice time is pivotal inmodeling confidence. We present a novel dualstage model that providesa unifying account for all four dependent variables, namely the collapsingconfidence boundary (CCB) model. According to CCB, decision isdetermined by a standard diffusion model. Choice, however, is followed bya second evidenceintegration stage towards a stochastic collapsingconfidence boundary
Keywords: Confidence, Diffusion model, Collapsing threshold/boundaries, Timepressure, Resolution of confidence, Choice, Perception.
DMB 2014 | 912 Sept 2014 | 27
Day 1 (Tuesday) 15h15Franc i s Cr ick Room
THOMAS HILLS University of WarwickTAKAO NOGUCHI University of Warwick
Choices can be risky. Unfortunately, more choices can be riskier. Wepresent research showing that in choices between monetary gambles,choice enriched environments increase risk taking and lead to less choicedeferral. Specifically, we compared decisions from experience withdecisions from description. In decisions from experience, when peoplesample a gamble they experience individual outcomes; people mustrepeatedly sample gambles to learn about the gamble's outcomes andprobabilities. In decisions from description, information about theprobabilities and outcomes are provided at the same time and repeatedsampling is unnecessary. A typical finding is that people underweight rareevents in decisions from experience and overweight rare events indecisions from description: a phenomenon called the descriptionexperience gap. To investigate the influence of set size on choice, weincreased the number of alternatives (to more than 30 options) in twostandard choice paradigms: The samplingparadigm (decision fromexperience), where people can sample freely over alternatives beforemaking a final decision, and the decision from description paradigm, wherepeople can deliberate over complete information about outcomes andprobabilities at the same time before choosing. In the first experiment,gambles were either very risky or fairly safe and payoffs were sampled fromsimilar distributions. Increasing set size increased the sample size in bothconditions, but led to fewer samples per gamble. In the sampling paradigmthis increased the likelihood of experiencing rare, risky events. When choicewas over gains, this led to increased choice for riskier options; peopletended to choose options associated with large (rare) gains. Over losses,set size had no influence on risky choice; encountering rare events ledparticipants to favor other risky gambles where rare events were notencountered, as they do for small set sizes. Importantly, the observationthat decisions from experience leads to underweighting of rare events wasreversed for large set sizes in the gains domain; instead, people chose as ifthey overweighted rare events. In a second experiment, we gaveparticipants £1 for showing up and then offered them the opportunity tobuy gambles with positive payoffs. We added a button to allowparticipants to, after free sampling of alternatives, either keep their money(defer choice) or to choose one of the alternatives. In both decisions fromdescription and decisions from experience choice deferral went down asset sizes increased. Moreover, in both experiments, choices were notchosen at random but were predicted by what the participants saw whenthey sampled alternatives. Overall, our results indicate that context effectsassociated with increasing set sizes are not always readily apparent fromtwoalternative decisions. However, the increase in risky choice associatedwith experienced gains is predicted from standard choice models. Theseeffects also extend to questions of choice deferral and informationoverload, for which we found no evidence of either.
Keywords: Risk, decisions from experience, too much choice, information overload,decisions from description
28 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 15h15James Watson Room
LIAM POLLOCK Queen Mary University of LondonDR KEITH JENSEN University of ManchesterDR MAGDA OSMAN Queen Mary University of London
In game theory the battle of the sexes (BOS) is a wellknown setupdesigned so that two agents (husband, wife) need to coordinate theiractions (go to the opera or the football) in order to maximise their payoffs.The game has two Nash equilibria (both act the same), each of whichfavours only one agent, or a mixed strategy (both act differently).
Typically, in experimental applications of the BOS, without signalling theiractions, adults follow inefficient strategies (poor coordination). Is this true ofchildren, particularly given how differently they conceptualise fairness? Toaddress this children aged between 513 were presented variations of theBOS game in which payoffs and signalling were manipulated, to seewhether reduced risk of opponent ’s outcome (condition 2) or reduceduncertainty of opponent ’s outcome (condition 3) would foster greatermutual cooperation as compared to the standard (condition 1).
304 children aged from 513 years (134 males, 170 females; mean age= 7.7 years, SD = 2.01) took part in a BOS game with 30 iterations. Childrenwere randomly allocated to one of three conditions: 1) standardBOS, 2)BOSincreased risk, and 3) BOSincreased risk + reduced uncertainty. Payoffmatrices for each of the 3 conditions are shown in Table 1.
In all conditions, children were tested in dyads facetoface across atable. Each player was given a stack of identical playing cards half ofwhich had lions on the face, the rest with foxes. Players were randomlyallocated an animal that was “ theirs” to indicate their preference. Themutual playing of “ their ” animal would maximise both the relative andabsolute payoffs for the animal ’s owner in the manner shown in table 1. Itwas clearly emphasised to all players during the habituation period that itwas their absolute and not their relative payoff that they needed tomaximise.
Our findings revealed that children showed the highest rates ofcoordination when uncertainty around the opponent ’s actions wasreduced, whereas for reduced risk and the standard conditionscoordination rates were comparably lower. Internal checks based onperformance in condition 1 and 2 suggest that our participants understoodthe experimental setup. In addition, we found that cooperation levelsbetween the three conditions were identical, meaning that our experimentalmanipulations impacted children’s decisions regarding cooperation, whichoccurred at a relatively low and, from a game theoretical perspective, is ahighly inefficient level. Implications for the ontogeny of decisionmakingrelated to cooperation and fairness are discussed.
Keywords: battle of the sexes, game theory, children, cooperation, coordination
DMB 2014 | 912 Sept 2014 | 29
Day 1 (Tuesday) 15h45Rosa l i nd Frank l i n Room
PETROC SUMNER Cardiff UniversityCRAIG HEDGE Cardiff UniversityALINE BOMPAS INSERM; CNRS; Lyon Neuroscience Research Center
Individual or group differences in rapid action decisions measured bytasks such as antisaccades and Stroop are generally assumed to reflectability to inhibit unwanted responses. However, there is little theoreticalunderstanding of what such a conclusion actually means and how togeneralize from it to realworld behaviour [13]. Progress is difficult becausedifferent response control tasks have typically shown surprisingly littlecorrelation (or little repeatability of correlation) with each other or with selfreported impulsivity. There is even little correlation between interchangeablyused measures from the same task – error rate (response control failure) andreaction time cost – which are universally assumed to tap the same process.This leads to confusion and apparent conflict in the literature.
Here, we replicate the lack of correlation between error and RT costsacross several response conflict tasks. We then show that this situation isstraightforwardly accounted for by competitive accumulator models inwhich impulsivity/reduced topdown control can be captured in (at least)two distinct ways: either as lower caution or lower selection ability (e.g., lessgood inhibition). Both of these increase errors, but they have oppositeeffects on RT costs. This means that if people vary in both caution andselection ability, there is expected to be no correlation between RT costand errors.
The reason caution and selection oppositely affect RT cost is thatincreasing caution trades fewer errors for increased RT costs (as well asoverall longer RT – the speedaccuracy tradeoff), while increasingselection bias reduces errors by limiting how much the ‘wrong choice’ getsactivated, and this reduces RT cost too. Therefore, although caution andinhibition/selection are both subsumed within the general concept ofimpulsive action, they are predicted to have opposite effects on somemeasures of response control. In this way the competitive accumulationframework makes an apparently confusing fact entirely understandable.
Keywords: Impulsivity; cognitive control; inhibition; speed accurary tradeoff.
30 | DMB 2014 | 912 Sept 2014
Day 1 (Tuesday) 15h45Franc i s Cr ick Room
JANINE CHRISTIN HOFFART University of BaselGILLES DUTILH University of BaselJÖRG RIESKAMP University of Basel
In recent years, there has been a growing interest in the socalleddescriptionexperience gap (DE gap) in risky decision making. The DEgap refers to the common finding that people choose differently in anexperiencebased choice paradigm and a descriptive choice paradigm. Inthe experienceparadigm people are allowed to draw samples to learnabout a gamble’s outcome distribution whereas in the description paradigmthey are presented with plain numerical descriptions of the gamble’soutcome distribution. The DE gap holds that in the description paradigm,people behave as if they overweighted relatively small probabilities,whereas in the sampling paradigm, people behave as if theyunderweighted small probabilities.
Many studies on the DE gap show that people draw relatively fewsamples before making a decision. Although this undersampling is oftenraised as an explanation for the DE gap, little is known about howprecisely sample size influences people’s choice behavior. A reason for thislack of understanding follows from the fact that successful models fordecisions from description, including prospect theory, make no predictionsabout sample size. Models that assume Bayesian updating, and to someextent memory based learning models, such as Gonzalez ’ instance basedlearning model (2003), do make predictions about the effect of samplesize. However, until now these effects have not been studied experimentallyyet.
Here, we apply a withinsubjects design to experimentally test the effectof sample size on choice behavior. For this purpose, we systematicallymanipulated the number of samples (5, 10, 20 or 40) that participants drewbefore making a choice. For each pair of gambles, the outcomes aparticipant saw matched the according gamble’s underlying outcomedistribution. In addition, the exact same gambles were presented in adescriptive format to the participants.
On an aggregate level, our data reveals that the DE gap is morepronounced for trials in which participants drew relatively small numbers ofsamples (i.e., 5 or 10 samples). So, it appears that overall, sample size doesinfluence people’s choices, yet it is not entirely clear to what extent. Onereason, for which we cannot draw strong conclusions about this influence, isthat binary choices contain little information about people’s degree ofpreference. Therefore, we conducted a second experiment, in which weadministered a gradual measure of people’s valuation of gambles. Such agradual measure allows us to apply more powerful statistical methods andthus to better understand differences in cognitive processes that might leadto the DE gap.
Keywords: Decisions from experience, sample size, DE gap
DMB 2014 | 912 Sept 2014 | 31
Day 1 (Tuesday) 15h45James Watson Room
PHILIP NEWALL University of StirlingBRADLEY LOVE University College London
A large body of empirical work in finance has concluded that mutualfund investors significantly reduce their welfare by selecting mutual fundswith high past returns rather than funds with low fees. Future performance ison average reduced oneforone with increases in fees. The large size ofthe mutual fund industry, and the dispersion of mutual fund fees, means thatthis is an area where behavior change interventions might increase investorwelfare. Potential interventions must, however, be evaluated for not onlytheir aggregate implications but potential perverse effects amongst subpopulations.
Previous work has attempted to debias mutual fund investors byreframing percentage fees into actual money amounts, where a 1% fee ona $10,000 portfolio would be reframed as $100 a year (Choi et al., 2010;Hastings & TejedaAshton, 2008). Although this effect has led to astatistically significant increase in feesensitivity, economically significantlevels of bias remain. Given the low cost of framing manipulations, this mightstill be a welfareincreasing intervention if investors respondhomogeneously.
In a simple 2x2 experiment with fees (actual money, percentage) andportfolio size ($1,000, or $1,000,000) manipulated betweenparticipants,we show perverse effects of actual money framing of fees: this manipulationleads to a large decrease in feesensitivity when the fee translates into alow amount of $10$15 a year, with the proportion of feeminimizingparticipants dropping from 40.6% to 27.4% (compared to a percentagecontrol). Furthermore, our results fail to replicate those of Choi et al. (2010)and Hastings and TejedaAshton (2008): representing the samepercentage fees as $10,000 or $15,000 led to a statistically insignificantimprovement over percentage framing (from 37.6% to 41.6%), despite percell averages of 250 participants.
Experiment two, which mirrored experiment one except that framing inpercentages or actual money was now applied to past returns, shows thatthis effect can instead induce greater feesensitivity. 54.9% of participantswho saw past returns of $10$15 a year minimized fees, a significantimprovement on 41.5% in the corresponding percentage group.
Our conclusions are twofold. Aggregate positive effects of anintervention can mask perverse responses amongst subpopulations.Interventions that improve behaviour on average may make certain groupssystematically worse. Second, we view our results as a simple proof ofconcept. Previous research has attempted to debias mutual fund investorsby increasing the salience of fees. Our results show that reducing thesalience of past returns may be a more effective method of increasing thedecision weight given to fees. Further work in experimental economicsshould aim to uncover efficient techniques of reducing the salience ofmutual fund past returns, as it shows promise as a method of debiasingindividuals in this economically significant market.
Keywords: Investing, Behavioural finance
32 | DMB 2014 | 912 Sept 2014
DMB 2014 | 912 Sept 2014 | 33
Day 1 (Tuesday) 16h45Rosa l i nd Frank l i n Room
ERICJAN WAGENMAKERS University of Amsterdam
People can learn through trial and error, a process that canbe quantified using models of reinforcement learning. Inpsychology, these models may be used for at least two purposes.First, researchers may be interested in what model best accountsfor people's choice behavior. This goal is usually referred to asmodel comparison or model selection. Second, researchers maybe interested in using a particular model to decomposeobserved choice behavior into distinct cognitive processes. Thisgoal is usually referred to as parameter estimation. Here I willdiscuss a research line that has provided a suite of tools formodel comparison and parameter estimation in the Iowa gamblingtask, one of the most popular reinforcement learning tasks to date.
ERICJANWAGENMAKERS
34 | DMB 2014 | 912 Sept 2014
DMB 2014 | 912 Sept 2014 | 35
Day 2 (Wednesday) 09h00Rosa l i nd Frank l i n Room
JEFFREY D. SCHALL E. Bronson Ingram Prof of Neuroscience, Vanderbilt University
The conjecture identifying spiking activity of certain neuronswith a stochastic accumulator decision process has inspired avigorous and productive research effort for 20 years. This efforthas described decision accumulation processes in multiple brainregions even extending to noninvasive ERP and fMRImeasurements. It has been buttressed by models of perceptualcategorization, response inhibition and visual search. Lately,though, several new findings have raised questions about thecoherence and clarity of the mapping between neurons anddecision accumulators. These include the first data showing howneurons identified with decision accumulators accomplishexecutive control and speedaccuracy tradeoffs and the firstmodel of response time from ensembles of accumulators. The newresults indicate that mapping between parameters of accumulatormodels and measurements of neural activity is not as transparentas originally presumed. The accumulator model framework will nodoubt remain an effective means of quantifying per formance andinstantiating computations in various tasks. However,the construction of a more secure bridge between model andneural levels of description will require more assiduityin (1) accounting for multiple stages of processing each addingtime and potential errors, (2) incorporating distinct neuralprocesses from heterogenous neurons in diverse neuralstructures, (3) articulating the transformations between spikes,ERPs, and BOLD, (4) specifying converging constraints to limitparameters in more complex models and (5) appreciating thelogical and rhetorical scope of the mapping — true identity,quantitative analogy, or interesting metaphor.
JEFFSCHALL
36 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
EUGENE MCSORLEY University of ReadingCLARE LYNE University of ReadingRACHEL MCCLOY University of Reading
It has been suggested that the evidence used to support a decision tomove our eyes and the confidence we have in that decision are derivedfrom a common source. Alternatively, confidence may be based on furtherpostdecisional processes. In three experiments we examined this. InExperiment 1, participants chose between two targets on the basis ofvarying levels of evidence (i.e., the direction of motion coherence in aRandomDotKinematogram). They indicated this choice by making asaccade to one of two targets and then indicated their confidence.Saccade trajectory deviation was taken as a measure of the inhibition ofthe nonselected target. We found that as evidence increased so didconfidence and deviations of saccade trajectory away from the nonselected target. However, a correlational analysis suggested they were notrelated. In Experiment 2 an option to optout of the choice was offered onsome trials if choice proved too difficult. In this way we isolated trials onwhich confidence in target selection was high (i.e., when the option to optout was available but not taken). Again saccade trajectory deviationswere found not to differ in relation to confidence. In Experiment 3 wedirectly manipulated confidence, such that participants had high or low taskconfidence. They showed no differences in saccade trajectory deviations.These results support postdecisional accounts of confidence: evidencesupporting the decision to move the eyes is reflected in saccade control,but the confidence that we have in that choice is subject to further postdecisional processes.
Keywords:Saccades DecisonMaking Confidence
Day 2 (Wednesday) 10h30
DMB 2014 | 912 Sept 2014 | 37
Day 2 (Wednesday) 10h30Franc i s Cr ick Room
NEIL BRAMLEY University College LondonDAVID LAGNADO University College LondonMAARTEN SPEEKENBRINK University College London
Most decision making research has focused on how people chooseactions in static contexts, where the structure of the task environment isassumed to be known and fixed. However, in real world decision making,people are typically uncertain about what model is true or appropriate forthe situation, meaning that part of the value of an action comes from theinformation its outcome can provide. The current research focuses on howpeople select sequences of actions (tests, manipulations or interventionson parts of a system) when their goal is to learn about the causal structureunderlying that system. Steyvers et al (2003) showed that people canoften select the most informative single variable to act on to distinguisheffectively between a set of alternative causal models in a partiallyprobabilistic scenario. We extend on this research over three experiments.
In experiment 1, we explored how people select and learn fromsequences of interventions on several fullyprobabilistic threevariablecausal systems, where they are allowed manipulate more than one variableat a time. We developed and tested models of participants' interventionchoices and causal structure judgments, using three measures of theusefulness of interventions: expected utility gain, probability gain andinformation gain. We also introduced plausible memory and processingconstraints. We found that successful participants were best described bya model that acts to maximize information (rather than expected score, orprobability of being correct); that discounts much of the evidence receivedin earlier trials; but that mitigates this by being conservative, preferringstructures consistent with earlier stated beliefs.
In experiment 2, we replicated our main findings from experiment 1 andestablished that providing summary information did not significantly reduceparticipants ’ accuracy. This suggested that limitations in memory forprevious interventions was not a key determiner of learning.
In experiment 3, we focused on identifying boundary conditions foreffective active causal learning and exploring active learning heuristics.We independently varied two aspects of the target causal systems: 1. Howreliable the causal connections were and 2. How frequently thecomponents of the system turned on by themselves. Within subjects, we alsovaried the number of causal components and number of connectionsbetween these components. We found that peoples ’ active causallearning abilities were robust and flexible across a wide range ofenvironments. Participants were able to identify causal connections evenwhen they were weak and unreliable but their learning broke down whenspontaneous activations were very frequent. Accuracy was unaffected bythe number of variables or links, but participants struggled with chainstructures in particular.
Overall, the picture emerges that human active learning uses a mixtureof heuristics. People exhibit periods in which they focus on “connecting” endorsing connections between actions and any apparent "effects",ignoring dependencies such as the existence of competing explanations.At other times they focus on “pruning” disambiguating causal paths andremoving unnecessary links from their existing model. We demonstrate that aheuristic active learning model following these principles can largelyexplain participants' performances.
Keywords: causal, learning, information theory, heuristics, active learning
38 | DMB 2014 | 912 Sept 2014
James Watson Room
FLORIAN ARTINGER Max Planck Institute for Human DevelopmentGORDON D. A. BROWN University of WarwickGRAHAM C. LOOMES University of Warwick
Individual preferences for risk are commonly elicited using Multiple PriceLists (Holt & Laury, 2002). How stable are individuals ’ preferences for risk asassessed by this method? Participants responded to each of four differentMultiple Price Lists on two separate occasions. Analysis of responses toidentical lists showed that participants exhibit imprecision in their revealedpreferences. Analysis of different Multiple Price Lists found that individuals ’preferences for risk systematically vary beyond the effect of imprecision. Thedifferences in CRRA obtained between different Multiple Price Lists arelarge and significant, with 34% of participants even switching from risk averseto risk seeking. The study highlights the context dependent nature of riskychoices in one prominent elicitation method and questions the reliability ofstable revealed preferences for risk.
Keywords: Risk preferences, context
Day 2 (Wednesday) 10h30
DMB 2014 | 912 Sept 2014 | 39
Rosa l i nd Frank l i n Room
ANDREAS JARVSTAD University of BristolIAIN D. GILCHRIST University of Bristol
The ability to inhibit taskirrelevant but salient stimuli, in favour of taskrelevant but less salient stimuli, is crucial for the cognitive control of action.Here, we explore this ability and two other key aspects of cognitive control(selection and choice), with a novel eye movement fMRI paradigm. Theparadigm avoids many of the potentially problematic confounds of existingtasks. It is continuous and naturally paced, involves looking at actualobjects, deciding which objects to look at and utilises saccadecontingentdisplays to equate visual stimulation across tasks. In contrast to previousstudies, we find that inhibition of taskirrelevant salient stimuli is associatedwith activity within the core saccadic network (posterior parietal cortex,frontal eye fields), seemingly without a strong reliance on areas external toit (e.g. dorsolateral prefrontal cortex, presupplementary motor area, inferiorfrontal gyrus). Our choice task, which attempts to expose the neuralsignature of choice, but contains many of the confounds present in existingtasks (e.g., task switching, task set complexity), activates the very sameextrasaccadic regions that have previously been linked to inhibition. Ourresults highlight the importance and challenges of cleanly isolatingcognitive functions and suggest limitations to the way cognitive functionsare mapped onto neural substrate.
Keywords: neuroimaging, motor control, eye movement, cognitive control, inhibition
Day 2 (Wednesday) 11h00
40 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
HRVOJE STOJIC Universitat Pompeu FabraPANTELIS P. ANALYTIS Max Planck Institute for Human DevelopmentHENRIK OLSSON University of Warwick
Predictions are essential to decision making. In making these predictionswe rely on knowledge of functional relations between attributes we canobserve and the outcome we seek to predict. A variety of tasks and modelshave been developed to investigate and explain how people learnfunctional relationships, but not in a decision making context. In a functionlearning task, participants learn to predict the continuous criterion of asingle alternative described by some attributes, while in a multiattributedecision making task, participants decide between two or more alternativesdescribed by some attributes. Differences between these two tasks might besubstantial enough that function learning operates differently in decisionmaking situations. The possibility of making comparative judgments in adecision making task, for example, is likely to change attention dynamics. Inaddition, because in a decision making task people receive feedback onlyabout the chosen alternative while at the same time unlabeled alternativesare present, the learning process might be better described by a semisupervised learning model.
Given the prevalence of decision making in the real world, much of ourconceptual knowledge is acquired while making decisions. Hence, our maincontribution of our results lies in establishing relations between two richliteratures function learning and decision making. We examine howlearning is affected by the decision making context and whetherqualitatively different models, rather than some of the classical functionlearning models, are needed to explain the learning process. We designedan experiment where a function learning task is yoked to a multiattributedecision making condition where participants also provide continuouspredictions. We have two function learning conditions where participantsare either yoked to see only alternatives that decision makers have chosen,or to see all the alternatives and receive feedback only on the chosenalternatives, but presented in separate trials.
Firstly, we find that decision makers learn functional relationships fasterand reach higher levels of accuracy than participants in the functionlearning tasks. Further, when we compare the decision making conditionwhere participants provide predictions with another decision makingcondition where participants do not make predictions, we find that both thedecision and learning performance deteriorates. To further investigate thesource of the differences between the conditions, we use a Kalmansmoother method to estimate the importance participants place on each ofthe attributes and fit formal learning models to the data. We find that asemisupervised form of learning is unlikely to be the source of theadvantage in decision making and further experiments are required tolocate the sources of the differences in learning.
Keywords: multiattribute decision making, function learning, learning, semisupervised learning
Day 2 (Wednesday) 11h00
DMB 2014 | 912 Sept 2014 | 41
James Watson Room
ANDREA POLONIOLI University of Edinburgh
Gigerenzer, Hertwig and their coworkers in the Centre for AdaptiveBehaviour and Cognition (ABC) and Centre for Adaptive Rationality (ARC)have recently articulated an innovative perspective on rational behaviorand cognition. I refer to this innovative framework as “adaptive rationality ”(henceforth, AR) for simplicity ’s sake. This framework encompasses adescriptive and a normative element. First, evolution has endowedorganisms with an “adaptive toolbox ” of simple heuristics, which are basicrules that are easy to apply. Second, although these heuristics are “ frugal”and may lead to violations of norms of the socalled “standard picture ofrationality ” (Stein 1996), they often lead to successful behavior and aretherefore “adaptively rational”.
This paper focuses on a challenge to this framework mounted byStanovich and his colleagues. Stanovich et al. (e.g. Stanovich and West2000) have conducted an important stream of individual differencesstudies involving reasoning and decisionmaking. An important result is thatnot everyone in reasoning and decisionmaking tasks follows “simpleheuristics”. Overall, there seems to be heterogeneity in the use of heuristics.Moreover, Stanovich et al. have found correlations between measures ofcognitive ability and tendency to follow standard norms of rationality.
I assess two arguments based on Stanovich’s research. The first argumentis supposed to challenge the adaptationist background of AR.Heterogeneity in the use of heuristics appears incompatible with the ideathat adaptationist pressures led to the use of heuristics: one would expectthe use of heuristics to be far closer to universal if they did. The secondargument seeks to question the normative claims made by AR theorists:according to Stanovich, the fact that people with higher cognitive abilitiesfollow standard norms of rationality suggests that the “standard picture ofrationality ” has normative force.
At least as things stand now, none of the arguments can be seen as fullycompelling. At least in principle, AR theorists can accommodate findings onheterogeneity, and I discuss several moves open to them. Moreover, eventhe most plausible version of the second argument turns out to beunconvincing. It is true that there is evidence that people with higher mentalabilities achieve better life outcomes (e.g. Deary et al. 2011), and onemight be tempted to conclude that they do so because they followstandard norms of rationality. This would suggest that, pace AR theorists,following standard norms of rationality does lead to successful behavior.Yet, the argument remains speculative, as we are unable to draw causalconclusions from the available data.
Whilst both of Stanovich's arguments miss their target, his research mightstill push forward the “ rationality debate” by putting forth two importantresearch questions and encouraging greater experimental and theoreticalwork. Following recent and seminal studies (e.g., McNamara et al. 2009;Mohlin 2012), AR theorists should offer concrete models and explain howevolutionary processes led to the heterogeneity in the use of strategiesthat is found in the lab. In addition, causally informative studies are neededto establish whether people with higher cognitive abilities achieve betteroutcomes because they follow standard norms of rationality.
Keywords: Heuristics, Evolutionary Psychology, Adaptive Individual Differences,Cognitive Ability, Cognitive Epidemiology
Day 2 (Wednesday) 11h00
42 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
ALINE BOMPAS Cardiff University; INSERM; CNRS; Lyon Neuroscience ResearchCenterPETROC SUMNER Cardiff University
Action decisions are considered an emergent property of competitiveresponse activations. As such, decision mechanisms are embedded in, andtherefore may differ between, different response modalities. Despite this, thesaccadic eye movement system is often promoted as a model for alldecisions, especially in the fields of electrophysiology and modelling, whileother research traditions predominantly use manual button presses, whichhave very different response distributions and are governed by differentbrain areas. Here we question and eventually validate the generalisabilityof core concepts of action decision models from saccadic to manualdomains, using two diagnostic phenomena: the remote distractor effect(RDE) and 'saccadic inhibition'. We find that, despite apparent differences,manual responses are also sensitive to the interference of visual distractorsand in fact the temporal profile of this interference indicates that the extradelay and variance in manual responses is best accounted for by nondecisional delays, not a difference in the decision process itself. Moreover,a biologicallyinspired model (DINASAUR, based on nonlinear inputdynamics) can account for both saccadic and manual responsedistributions by simply adjusting the balance of fast transient and slowsustained input.
Keywords: Action selection; response modalities; competition; reaction time
Day 2 (Wednesday) 11h30
DMB 2014 | 912 Sept 2014 | 43
Franc i s Cr ick Room
ALICE MASON University of BristolCASIMIR LUDWIG University of BristolPAUL HOWARDJONES University of BristolSIMON FARRELL University of Bristol
The release of dopamine, in response to rewarding and motivationallysalient events, is thought to increase plasticity in the hippocampus. Thisresults in enhanced memory encoding for rewarding events (Lisman &Grace, 2005). How best to promote learning is a central question for bothresearchers and educators. In a series of studies we asked whetheruncertainty of reward outcome has an effect on rewarded memory. Weused a simple verbal memory task experiment in which participants weretold they could earn small monetary incentives for each word theysuccessfully remembered. Some of the rewards were fixed and others weredelivered with a 50:50 probability of reward or no reward. Twenty four hoursafter the learning phase participants were invited back to lab to perform arecognition memory test. In Experiment 1, we tested whether uncertainrewards promote learning to a greater extent than certain rewards. Basedon observations of a sustained ramping response in striatal dopaminebetween the cue signally reward and reward delivery (Fiorillo, Tobler &Schultz, 2003), we predicted that learning would be greatest duringuncertain trials, when there is maximum dopamine for hippocampaldependent consolidation. We found uncertain reward produced a memoryenhancement similar in magnitude to fixed reward. In Experiment 2, we useduncertain and certain reward cues, but we also revealed the rewardoutcome for each trial during the learning phase, so that the design wasanalogous to the neurophysiological studies. Mixed effects modelling of thedata revealed that uncertainty and prediction error are significant driversof recognition memory. Lastly, we wanted to investigate if uncertain rewardslead to better memory encoding than certain rewards regardless of value.We designed an experiment with uncertain and certain reward values andfound superior recognition memory under uncertainty. In our last experimentwe have been looking at the differences immediate and delayed memoryrecall. If these effects are dependent upon increased hippocampalconsolidation we would expect them to only be evident or significantlyincreased after a twenty four hour delay. These studies are the first todemonstrate an influence of uncertainty of reward on motivated learning.Chance based uncertainty is seen as a key component of games and is ofcore interest to educators in understanding how best to maximise rewardbased learning (HowardJones et al, 2011).
Keywords: Reward Uncertainty, Incentivised learning, Episodic memory
Day 2 (Wednesday) 11h30
44 | DMB 2014 | 912 Sept 2014
James Watson Room
GRAHAM LOOMES Warwick Business SchoolGANNA POGREBNA Warwick Manufacturing Group
Subjective Expected Utility (Savage, 1954) provides a normativeaccount of decision making under risk (when probabilities are known) andunder ambiguity (when probabilities are not objectively known). SEUTpostulates that individuals have neutral attitude towards ambiguity andsubjective probabilities of ambiguous events could be extracted frompreferences (Machina and Schmeidler, 1992). Yet, the Ellsberg paradox(Ellsberg, 1961) questions the descriptive validity of SEUT by providing anexample where individuals exhibit nonneutral attitude towards ambiguity.
To date, the validity of SEUT and other theories of decision makingunder ambiguity have been primarily tested in games of nature where anindividual faces ambiguity which is constructed using a noninteractiveenvironment. However, in the majority of the real world situations ambiguityarises from the difficulty of predicting the actions of others. This paperexamines the implications of SEUT and other theories of decision makingunder ambiguity in games of strategy where individuals face ambiguitywhich stems from an interactive environment.
We design a simple experiment where participants are faced with aclosed set of three items. Participants are informed that in order to receivea monetary payoff of £20, each of them needs to coordinate with anotherrandomly selected participant by choosing the same item from the closedset. Failure to coordinate results in a zero payoff. Then participants areasked to think about the likelihood of the different choices the otherparticipant could have made. Such events are ambiguous because theydepend on the actions of other people and the probabilities are notobjectively known.
We elicited certainty equivalents and probability equivalents for eachambiguous event separately as well as for pair of events within the closedset. We compare certainty and probability equivalents obtained fromdifferent decision problems and analyse whether and to what extentindividuals make consistent decisions throughout the experiment andwhether these decisions can be accounted for using SEUT and otherdeterministic theories. We find that individuals violate SEUT in this strategicenvironment. In particular, in one subset of experimental tasks manyparticipants exhibit ambiguity averse behaviour, while in another subset theirbehaviour is often consistent with ambiguity seeking. We provide anexplanation of observed behaviour which is based on a model ofstochastic choice.
Day 2 (Wednesday) 11h30
DMB 2014 | 912 Sept 2014 | 45
Rosa l i nd Frank l i n Room
PHILIP PÄRNAMETS Lund University Cognitive Science
Eye gaze reflects online processing in many cognitive tasks, such asspatial memory, and linguistic processing (Spivey, 2007). Choices, as theyunfold over time, are also reflected in concurrent activation of perceptualand motor systems (Gold & Shadlen, 2000; McKinstry, Dale, Spivey, 2008).This allows for both faster action implementation and for the cognitivesystem to receive direct feedback from the environment. However, effects ofgaze distribution on higherlevel, abstract choices, such as moral ones,remain unexplored. Previous studies have established how gaze can affectchoice, but only for simple preferential choices, and only by directlycontrolling the information that is attended to (Shimojo et al. 2003; Armel etal. 2008). Here we investigated this aspect further; we show that knowledgeof gaze dynamics can be leveraged to influence choices, even thoseconcerning complex moral issues, by manipulating the timing of the choicealone.
We monitored participants ’ eye movements during a twoalternativeforced choice task with moral and nonmoral factual questions. Participantswould hear a statement read aloud and see two response alternatives. Forexample, they would hear a statement such as “Murder is sometimesjustifiable” and see the alternatives “Sometimes justifiable” and “Neverjustifiable”. One alternative was randomly predetermined as a target for themanipulation. At the moment participants had fixated the target alternativefor at least 750ms and the nontarget for at least 250ms, their deliberationwas terminated and their choice was prompted. The timing and cause ofthe choice prompts was not known to the participants nor discovered bythem during the course of the experiments.
We found that participants choose the target alternative in 58% of thetrials for the moral statements and 55% of the trials for the factual nonmoralstatements. Response times from choice prompt to button press weresignificantly faster when choosing the target alternative as was participantconfidence.
Having demonstrated this causal relationship between eye gaze andmoral choice we also attempted to better characterise the underlyingcomputational mechanisms. We fit diffusion models (Krajbich & Rangel,2010) with and without dependence on eye gaze to data on moralchoices using the same statements as before. We found that an eye gazedependent model provides a better fit and can capture many features ofthe empirical data.
We conclude that by tracking the interplay between a decision maker,her perceptualmotor systems, and the environment, it is possible toinfluence the outcome of a choice without manipulating the informationavailable to her. Highlevel, abstract cognition, such as moral choices, ispartly caused by interactions with the immediate visual environment, andthis can be described using similar models as for simple preferential andperceptual choices.
Keywords: Moral choice, visual attention, modelling
Day 2 (Wednesday) 12h00
46 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
TIM W. FAWCETT University of BristolBENJA FALLENSTEIN University of BristolANDREW D. HIGGINSON University of BristolALASDAIR I. HOUSTON University of BristolDAVE E.W. MALLPRESS University of BristolPETE C. TRIMMER University of BristolJOHN M. MCNAMARA University of Bristol
Patterns of decisionmaking in humans reveal some striking deviationsfrom economically rational expectations. These include distorted beliefsabout external events, inconsistent preferences that are altered by pastexperience and current context, and apparent violations of the axioms ofrational choice theory. There is mounting evidence that analogous biasesexist in other organisms, which hints at an evolutionary explanation. Yet suchcases of apparently irrational behaviour seem difficult to reconcile with thefundamental biological concept of natural selection as an optimisingprocess. Here I summarise recent theoretical research from our group thatsuggests a possible solution to this conundrum. Our work demonstrates thatseveral common biases in decisionmaking may in fact result fromecologically rational decision rules adapted to exploit temporallyautocorrelated environments. Temporal autocorrelation is ubiquitous innature and implies that conditions experienced now or in the recent pastare predictive of future conditions, which can affect the consequences ofcurrent decisions. Many standard laboratory procedures used to studydecisionmaking do not reflect this statistical structure, and in such settingsecologically rational decision rules can lead to biased or irrationalbehaviour. This evolutionary perspective can account for some wellknowndeviations from economic rationality, including intransitivity and irregularity,successive contrast effects and the fourfold pattern of prospect theory. Weencourage other researchers to consider the richness of naturalenvironments to understand better how evolution has shaped our cognitivesystems. The real world can be complex, variable and autocorrelated, andwe should expect cognitive and perceptual systems to have evolved toexploit its statistical structure.
Keywords: cognitive bias, irrational behaviour, evolution, economic rationality,ecological rationality, temporal autocorrelation, intransitivity, irregularity, successivecontrast effect, prospect theory, fourfold pattern
Day 2 (Wednesday) 12h00
DMB 2014 | 912 Sept 2014 | 47
James Watson Room
DAVID KELSEY University of ExeterSARA LEROUX Oxford Brookes
We report an experimental test of the influence of ambiguity onbehaviour in a coordination game. We study the behaviour of subjects inthe presence of ambiguity and attempt to determine whether they prefer tochoose an ambiguity safe option. We find that this strategy, which is notplayed in either Nash equilibrium or iterated dominance equilibrium, isindeed chosen quite frequently. This provides evidence that ambiguityaversion influences behaviour in games. While the behaviour of the RowPlayer is consistent with randomising between her strategies, the ColumnPlayer shows a marked preference for avoiding ambiguity and choosing hisambiguitysafe strategy.
Keywords: Ambiguity; Choquet expected utility; coordination game; Ellsberg urn,experimental economics
Day 2 (Wednesday) 12h00
48 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
JIAXIANG ZHANG Medical Research Council, Cambridge, UK
The search for symptomatic and diseasemodifying therapies forneurodegenerative diseases urgently requires a mechanistic understandingof cognitive deficits in patients. Here we present two recent studies ondecisionmaking impairments in different cognitive domains.
The first study investigated impairments of response inhibition inProgressive Supranuclear Palsy (PSP) and Parkinson’s disease (PD). The twodiseases have different neuropathology, but both affect cognitive functionincluding impulsivity. We tested nineteen PSP, twentyfour PD and twentysixhealthy controls in a saccadic go/nogo task by using a headmountedinfrared saccadometer. Subjects were cued on each trial to make a prosaccade to a horizontal target or withhold their saccadic responses. Wemodelled the paradigm as a diffusion process between a “go” and a “nogo” decision boundary and fitted a hierarchical driftdiffusion model tobehavioural data by using Bayesian parameter estimation approach. BothPSP and PD patients were impulsive: they had more commission errors in the“nogo” condition, but the commission errors and mean saccade latencieswere similar between PSP and PD patients. However, the posterior modelparameters revealed significant diseasespecific deficits in the mechanismsunderlying go/nogo decisions. Both PSP and PD patients had slowerinformation accumulation rate and faster nondecision time than controls,but PSP patients were severely biased towards the prosaccade decisionboundary than PDs and controls. The combination of a welltoleratedoccolmotor paradigm and the sensitivity of the Bayesian modellingapproach provide a useful biomarker for distinguishing differentneurodegenerative diseases and provide a rational basis on which todevelop and assess new therapeutic strategies.
In the second study, we examined the performance of subsecondinterval discrimination in PD. Eighteen nondemented and medicated PDpatients and nineteen healthy controls performed two durationdiscrimination tasks, during which the participants were required tocategorize auditory durations into two (bisection) or three (trisection)interval categories. The patients with PD exhibited impaired intervalsensitivity than the controls in both tasks and biased interval decisions inthe bisection task. The PD patients also showed maladaptive decisionstrategies that failed to slow down in discriminating ambiguous intervals.Furthermore, the impaired timing performance in the PD patients correlatedwith their disease severity and doses of dopaminergic medication. Thefindings suggest that impaired interval discrimination in PD cannot simply beattributed to pathophysiological distortions in an internal clock, but alsoassociated with impulsive decisionmaking processes which are biasedtowards premature responses.
Keywords: Decisionmaking, go/nogo, interval discrimination, PD, PSP, modelling
Day 2 (Wednesday) 14h00
DMB 2014 | 912 Sept 2014 | 49
Franc i s Cr ick Room
GORDON D. A. BROWN University of Warwick
How do people value states of health, decide on a fair price for aproduct, or determine the appropriate amount of damages to awardagainst a polluting company? Here I describe and discuss a process I call“ relative rank matching”. The subjective magnitude of quantities such asprices, health states, or crimes are assumed to be determined by contextualcomparison involving rankbased principles such as those embodied inRange Frequency Theory and Decision by Sampling. However, such modelsare often silent on the question of how comparisons are made acrossincommensurable dimensions. Subjective judgements are assumed to beentirely relative, yet we have no difficulty rejecting a “ relatively good”bottle of wine in favour of a “ relatively bad” house. Although relativejudgements cannot themselves provide a common currency for comparingoptions, it is suggested that relativitymatching translation into a commondistribution (such as market prices) is often possible and enablescomparison across different dimensions. When a suitable matchabledimension such as a market price distribution is unavailable, however, ourvaluations are inconsistent and unreliable.
Keywords: Judgement; decision making; relative rank
Day 2 (Wednesday) 14h00
50 | DMB 2014 | 912 Sept 2014
James Watson Room
ULLRICH ECKER, University of Western AustraliaSTEPHAN LEWANDOWSKY, University of Bristol
The quality of decisions is determined largely by the quality of theinformation they are based on. Misinformation and inaccurate beliefs aretherefore of major concern: If a majority believes in something that isfactually incorrect, the misinformation may form the basis for political andsocietal decisions that run counter to a society ’s best interest; if individualsare misinformed, they may likewise make decisions for themselves and theirfamilies that are not in their best interest and can have seriousconsequences. For example, following the unsubstantiated claims of avaccinationautism link, many parents decided not to immunize theirchildren, which has had dire consequences for both individuals andsocieties, including a marked increase in vaccinepreventable disease andhence preventable hospitalizations and deaths, as well as unnecessaryexpenditure for followup research and publicinformation campaigns aimedat rectifying the situation.
Misinformation comes in many guises, including explicitly providedfalsehoods, subtly misleading headlines, or urban myths. What makesmisinformation particularly concerning is its resistance to correction. In thistalk, we will illustrate how misinformation affects memory, reasoning, anddecision making, and we will present experimental evidence for theineffectiveness of corrections. We will discuss the role of people’s attitudestowards the corrected misinformation, and will highlight strategies toincrease the potency of myth rebuttals and refutations.
Keywords: Misinformation; Reasoning; Memory
Day 2 (Wednesday) 14h00
DMB 2014 | 912 Sept 2014 | 51
Rosa l i nd Frank l i n Room
ENCARNI MARCOS Universitat Pompeu Fabra, BarcelonaPIERPAOLO PANI Sapienza University of RomeEMILIANO BRUNAMONTI Sapienza University of RomeGUSTAVO DECO Universitat Pompeu Fabra, Barcelona; ICREA InstitucióCatalana de Recerca i Estudis Avançats, BarcelonaSTEFANO FERRAINA Sapienza University of RomePAUL VERSCHURE Universitat Pompeu Fabra, Barcelona; ICREA InstitucióCatalana de Recerca i Estudis Avançats, Barcelona
In decision making information stemming from both memory andperception is integrated in the service of the goal oriented action of theagent. In a previous study, we have proposed that the behavior of theagent is biased by the sequentiality of external events and that this biasingis achieved by a proper chaining of the memory space (Verschure et al.,2003). However, the underlying neural mechanisms causing this bias inbehavior are not well understood yet. Here, we investigate this issue bylooking at the activity of neurons in the dorsal premotor area (PMd) of twomacaque monkeys while they performed a countermanding reaching task.The countermanding task has been extensively used to study the ability tocancel a planned movement when an infrequent delayed stop signalappears. In this task, the probability of success mainly depends on thetemporal distance between a gosignal (a visual cue that indicates that amovement should be initiated) and the socalled stopsignal (a secondvisual cue that in a minority of trials indicates that the previous plannedmovement has to be cancelled). Recent results have shown that thebehavior of subjects in a single trial is also influenced by the history of thattrial, i. e. reaction time (RT) of subjects is longer when one or many Stoptrials have been recently experienced (Emeric et al., 2007). In the presentstudy, we investigate the neural mechanisms causing this modulation inbehavior due to the task history. We found no relationship between meanfiring rate of PMd neurons and the changes of RT due to previousexperience, i. e. whether a Go trial was preceded by a Stop or a Go trial.In contrast, we observed that the acrosstrial variability of the neuralresponses (measured by the variance of the conditional expectation,VarCE, Churchland et al., 2011) showed the same modulation with trialhistory than the mean and STD of the RT and that this variability could beused as a predictor of RTs in these cases. An additional test with atheoretical model suggested that a system that continuously monitors therecent history of a trial is necessary and sufficient to explain the observedbehavioral and neural modulation with trial history. These results raise twoimportant questions: whether this monitoring system actually exists andwhether it lies inside or outside PMd.
Keywords: decisionmaking, variance, trial history, PMd, motoring system
Day 2 (Wednesday) 14h30
52 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
TATIANA KORNIENKO University of Edinburgh
How happy would Robinson Crusoe be with a particular coconut on hisdesert island? Undoubtedly, that would depend on the coconut's size. Yet,measuring tapes and scales are evolutionarily advanced technology,inaccessible until very recently. How would Nature enable individuals tojudge, or measure, magnitudes, and what is the relationship of suchmagnitude judgments to a utility function?
I explore the process and tools which can be used to judge, or measure,a given magnitude. Following Robson [2001, 2002], I assume that, ratherthan endowing living organisms with all necessary judgment scales, Natureonly provides "tools" that enable one to extract information from one'senvironment and experience. To deal with more evolutionarily recent tasks,the candidate cognitive tools have to be independent of the nature of themeasurement.
I identify two welldocumented domainindependent cognitive toolsand, using a parsimonious mathematical model, show that one can judge, ormeasure, a magnitude of an item entirely using ordinal comparisons which,by means of a frequency (proportion) tool, are keyed into a universalcardinal scale. By calculating how frequently a given object "wins" apairwise ordinal "tournament" against all other objects in the reference set,one can map an arbitrary set of modalities (such as quantities, sizes,weights, durations, luminosities, and so on) onto the interval [0,1]. Suchprocess of evaluation by ordinal rank is a basis of "decision by sampling" inStewart, Chater and Brown [2006], which been used to explain "preferenceanomalies".
I argue that the resulting magnitude judgment is a candidate for a utilityfunction representing "more is better" preferences. That is, a utility functionmay simply be an artefact of magnitude judgment. Yet such magnitudejudgment is referencedependent, leading to a possibility that one'smarginal utility (and thus apparent attitudes towards risk) could be shapedentirely by one's experiences, memory, and cognitive imperfections. I showthat a (neoclassical) contextindependent utility function may arise as aspecial case of this magnitude judgment procedure when one judges amagnitude relatively to a remembered sample and has a long memory. Incontrast, if one's memory is bounded, the well documented context effectstend to arise.
The model presented here allows to accommodate well knownperceptual errors in the two cognitive tools and to explore the effects ofsuch cognitive limitations on magnitude judgment. One notable result is thata certain class of cognitive distortions of the mental line would lead toapparent risk aversion in an environment where risk neutrality would havebeen optimal in the sense of Robson [2001] and Netzer [2009].
Despite the asocial nature of cognitive processes, the procedureinvolves builtin social comparisons. I show that welfare evaluation ofredistributive policies and economic growth depend on the bias in ordinalimperfections. Furthermore, as judgment of own skill relatively to the othersinvolves interpersonal comparisons, the proposed model allows one torelate individual difficulties in making upward ordinal comparisons with theobserved patterns of overconfidence in relative skill judgment.
Keywords: Referencedependent preferences, random utility, procedural approach,memory limitations, perceptual imperfections, WeberFechner law, decisionbysampling, overconfidence, welfare.
Day 2 (Wednesday) 14h30
DMB 2014 | 912 Sept 2014 | 53
James Watson Room
SUDEEP BHATIA University of WarwickNEIL STEWART University of Warwick
Multiattribute choice is an important area of enquiry in decision makingresearch. While existing work on multiattribute choice has greatly improvedour understanding of behavior, it suffers from an important limitation. Multiattribute choice theories need observable and quantifiable attributes inorder to be tested. However, most everyday choices from deciding whichmovie to watch, to what food to eat involve objects defined on latentattribute dimensions. How can we study everyday multiattribute choicewithout quantified, observable information about the attributes that thesechoices involve?
In this paper we offer a solution to this problem. We argue that largedatasets with detailed text descriptions of everyday choice objects canbe used to uncover the latent attributes that these objects are defined on.These latent attributes can then be used to study multiattribute choice inits naturalistic setting.
We applied our approach to predicting movie choices, using text datafrom the Internet Movie Data Base (IMDB). Particularly, we obtained the plotsynopses of the 500 most voted movies on IMDB. These synopses had anaverage length of 1,025 words, and offered detailed descriptions of themovie plots of the most popular movies in the world. We processed thesesynopses with Latent Semantic Analysis (LSA), a natural languageprocessing technique, used to analyze the conceptual structure of textcorpuses. By applying LSA to our dataset, we were able to discover thelatent attribute dimensions that characterize a decision maker ’s movieuniverse.
Having obtained quantified attribute values for different movies, wetested whether these attributes did in fact play a role in decision making.For this, we asked 75 participants to make 200 hypothetical choicesbetween different movie triplets, selected from the 100 most voted movieson Netflix. We assumed that decision makers utilized a probabilistic versionof weightedadditive decision rule, applied to the top five latent attributesobtained from LSA. We fit both grouplevel and individuallevel attributeweights to the movie choices obtained in our experiment, and found thatour model provided a good fit to the data. Particularly, 55 out of the 75participants had a statistically significant fit (average loglikelihood =207.99, p<0.05 using the likelihoodratio test) with the weighted additivemodel on the five latent attributes. Our model also had a statisticallysignificant fit on the pooled choice data from our experiment (loglikelihood = 16,322.96, p<0.01 using the likelihoodratio test).
Overall the result of our study show that latent attributes obtained fromlarge text corpuses can be used to predict everyday choice. Decisionmaking research does not need to limit itself to artificial decisionenvironments in which attribute information is explicitly presented toparticipants as part of an experimental choice task. Rather, online datasetscombined with powerful text processing techniques, can extend the insightsof multiattribute decision research to the naturalistic settings in which mostof our choices are made. Future work will extend our analysis to othernaturalistic domains, and additionally use these domains to compare thedescriptive power of different behavioral theories of multiattribute choice.
Keywords: decision making, latent semantic analysis, multiattribute choice
Day 2 (Wednesday) 14h30
54 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
GILLES DE HOLLANDER University of Amsterdam, Amsterdam Brain and CognitionANDREAS SCHÄFER University of Amsterdam, Amsterdam Brain and CognitionROBERT TURNER University of Amsterdam, Amsterdam Brain and CognitionBIRTE FORSTMANN University of Amsterdam, Amsterdam Brain and Cognition
The subthalamic nucleus (STN) is a nucleus in the basal ganglia ofapproximately 100mm3 in size and a crucial node in the corticobasalganglia decisionmaking network. It is theorized to act as a global brake onthe decision process, heightening the threshold during difficult decisions, aswell as implementing stopping behavior.
Parkinson's disease (PD) is charachterized by the loss of dopamenergiccell in the substantia nigra and, consequently, a hyperactivated STN. PD isnow often treated using deep brain stimulation (DBS) of the STN, alleviatingits 'motor' symptoms such as tremor. This treatment is known to induce severesideeffects such as cognitive decline, compulsive gambling, depression,and even suicide. In the clinical literature it is often proposed that the STNcan be subdivided in discrete cognitive, limbic and motoric parts,differentially connected to functional counterparts in the cortex (Temel etal. 2005). The sideeffects of DBS are then explained by the stimulation ofnonmotoric parts as opposed to the motor part of the STN. A recentreview of the empirical literature (Keuken et al, 2012), however, showsincosistent results in the number and location of these subdivisions acrossstudies.
In this study we aimed to investigate these putative subdivisions usingultrahigh resolution 7T functional magnetic resonance imaging (fMRI), witha voxel resolution of 0.83x0.83.1.0 mm and individually segmented STNmasks, which is much more detailed approach than the oftenusedresolution at 3T imaging of approximately 3mm isotropic with standardpopulation masks/coordinates. We used a newly developed decisionmaking paradigm, which contained 3 experimental manipulations, targetedat inducing differential activation of these putative subdivisions and theircortical counterparts.
The subjects were shown pictures from the IAPSdataset, a standarizeddataset of emotionally valenced pictures. One half of these pictures wererated as of neutral emotional valence, the other half as of negativeemotional valence. The pictures were colorized using a filter, making themslightly green or slightly red. One half of the pictures was more colorizedthan the other half. The goal of the subject was to indicate whether thepicture they saw was red or green. Additionally, on 25% of the trials anauditory stopsignal was presented and the subject was instructed to thenwithold their response.
There were four experimental conditions that are related to fMRI BOLDsignal in both cortex and STN:1. Emotional Valence (limbic network) 2.Difficulty (cognitive network) 3. Stop Trial (cognitive/motor network) 4.Response direction (motor network). The stop signal linear ballisticaccumulator model was used to control for any interactions betweenemotional valence, difficulty and stopping behavior.
Preliminary, unpublished results, both behaviorally and from theneuroimaging will be presented and discussed.
Keywords:7T fMRI, Basal Ganglia, Subthalamic Nucleus, Subdivisions, Topography
Day 2 (Wednesday) 15h00
DMB 2014 | 912 Sept 2014 | 55
Franc i s Cr ick Room
SARAH SMITH University College LondonADAM J. L. HARRIS University College London
Evidence that preferences may be relative not absolute raises concernsabout the practice of using risk attitude measures to guide investmentchoice. For example, Stewart et al. (2003) found that valuations of gamblesalmost entirely depended on the set of choices presented, and Vlaev et al.(2007) found similar effects in the selection of pension funds. In the currentstudies, we investigated the relativity of risk attitude by manipulating theprior experience of participants, rather than the options available to themin the immediate choice context. This addresses the potential that previousfindings arose because the constrained set of choices did not include thedecision maker's (absolutely) preferred option, and findings reflect a forcedchange in choice strategy rather than a lack of absolute preference.Additionally we aimed to test the degree to which the effects are the resultof a judgment system that is fundamentally relative, rather than an easilyoverridden heuristic.
We conducted two studies in which the experimental manipulation wasa prior experience of making pairwise choices between pension funds, andthe dependent variable was the final choice of a fund from a full set. Thus, ifa decisionmaker had an internal, stable, absolute risk preference shecould make her final choice in line with it. Study 1 had four conditions: priorexperience constrained to high risk, to low risk, or to balanced prospects,and no prior experience. We found condition significantly affected finalfund choice. Participants whose prior experience was of high risk prospectsmade a higher risk final choice from the full range. However, there was alsoevidence of some sensitivity to absolute values.
Study 2 examined context effects withinparticipants and tested a biaswarning to see if context effects result from a heuristic process that can beoverridden. All participants experienced either the high or low risk pairwisechoices and then made a choice from the full range. Fourteendays laterthey did the opposite test. Half of the participants were warned about thepotential influence on them of the choice set. We found significant contexteffects. Participants selected a lower risk fund after experiencing low riskprospects than they did after experiencing high risk prospects. Effects werenot influenced by order, or attenuated by the bias warning. Again therewas evidence of absolute preference: at the participant level 29% ofparticipants chose the same investment fund after both high and low riskexperiences.
These prospect relativity effects violate economic principles of stablepreference and are similar to those found in psychophysical judgments.They are consistent with Range Frequency Theory (Parducci, 1965), andwith a Decision by Sampling account of the choice process (Stewart et al.2006). However, absolute preference appears to also have a role. Howabsolute preference and relative judgments might work simultaneously toinfluence risk attitude needs consideration. From an applied perspective,customers' risk preferences are susceptible to manipulation in the immediatedecision making environment, and in any preceding preparation tasks oronline journeys.
Keywords:judgment and decision making, prospect relativity, decision by sampling,financial risk attitude
Day 2 (Wednesday) 15h00
56 | DMB 2014 | 912 Sept 2014
James Watson Room
JULIA BRETTSCHNEIDER University of WarwickANJALI MAZUMDER University of Warwick
One of the recent developments in medical genomics are refinedprognostic tests providing socalled recurrence scores for common cancerssuch as breast, prostate and colon to be used for individualized decisionmaking about adjuvant treatment. They add an additional piece ofinformation to an already complex decision process involving clinicalevidence such as cancer stage, eligibility for treatment, health, life style andpersonal preferences. Such decisions involve a delicate tradeoff betweenthe shortterm health risks and decreased quality of life on the one handand potential decrease of total years of survival on the other hand.
In this project we characterize and define adjuvant treatment decisions,a class of decision tasks that includes the adjuvant treatment examplesdescribed above, model the decision making process with a particularemphasis on deviations from normative rules of probability and propose aprototype for a support tool.
Such genomic recurrence tests are cover by some US health insurancespolicies and by the Canadian health care system. The recurrence scoresare numerical predictions for the recurrence risk based on the expressionvalues of a panel of genes in the patient ’s tissue sample. (For example, forbreast cancer, the American Oncotype DX test is based on a 21genesignature measured by RTPCR, and the Dutch MammaPrint test is based ona 70gene signature assessed by a custom microarray. The predictive valueof these tests has been established in initial studies, and longterm clinicaltrials are currently under way.)
Particular characteristics of adjuvant treatment decision problems areuncertainty and ambiguity on a number of levels, including reliability of thetest results, validity of the genomic recurrence score and contradictions withtraditional clinical information. In addition, they are not transparent, butbuild on recent developments in genomics, which not only remain a blackbox for patients and health care professionals, but are not yet fullyunderstood scientifically. Furthermore, attitudes towards genomics as a newtechnology play into the decision process. In addition, mainly for patients,strong emotions such as fear, shame or anger can interfere with theirdecision process. Furthermore, the patient ’s social support system can playa central role.
We develop a rational model using Bayesian networks for decision tasksof our adjuvant treatment decision class and develop scenarios to gaugethe impact of the modifying factors described above. We put a particularemphasis on the processing of risk information, such as an investigation ofhow fallacies known from descriptive studies may potentially occur in ourdecision class. Using a prospect theory framework, we analyse the effect ofdifferent risk attitudes. Different preferences are also studied and can varybetween physicians and patients. In a context of shared decisionmakingthis can be used to detect and overcome hidden obstacles. Finally, wepropose a prototype for a decisionsupport tool that supports patientsand physicians while they go through a complex decision task, with the aimof making this complex process transparent and manageable, whileavoiding unintended fallacies.
Keywords: cancer recurrence risk, adjuvant treatment decision, decision biasfallacies, decision support
Day 2 (Wednesday) 15h00
DMB 2014 | 912 Sept 2014 | 57
Rosa l i nd Frank l i n Room
NICK CHATER Professor of Behavioural Science, Warwick Business School
Many important decisions involve interacting with otherdecision makers. What I should do depends on what other peoplewill do. But this can lead to some apparently awkward problems ofcircularity. So person A can only decide what to do on the basisof what person B will do; but B person B can only decide to doon the basis of what person A will do! One way out of thisdeadline is the notion of Nash equilibrium, as widely applied ineconomics. Here I explore an alternative way of breaking thedeadlock, virtual bargaining, developed with Jennifer Misyak,Tigran Melkonyan and Hossam Zeitoun. Virtual bargainingproposes that decision makers implicitly compute what they wouldagree to do, if they were to bargain explicitly; and then carr y outthat agreement. I argue that virtual bargaining may underpinsocial behaviour and communication, and perhaps bedistinctively human.
Day 2 (Wednesday) 16h45
NICKCHATER
58 | DMB 2014 | 912 Sept 2014
DMB 2014 | 912 Sept 2014 | 59
Rosa l i nd Frank l i n Room
TOM GRIFFITHS Associate Professor of Psychology and Cognitive Science,University of California, Berkeley
Bayesian models of cognition explain human inductive leapsas probabilistic inferences, combining prior knowledge withavailable data. This raises two important questions: how can weidentify the prior knowledge that people use, and what cognitiveprocesses make it possible for people to act consistently withBayesian inference despite its computational intractability? I willpresent possible answers to both questions, describing anexperimental framework for estimating prior distributions that ismore effective than traditional elicitation methods and atheoretical framework for connecting Bayesian models ofcognition to psychological process models and heuristics. Thistheoretical framework which we call "resource rationality" assumes that people make effective use of their limited cognitiveresources. I will argue that this approach gives us formal tools fordefining and deriving heuristics, and show that classic heuristicsfrom the decisionmaking literature naturally fall out of this kind ofaccount.
Day 3 (Thursday) 09h00
TOMGRIFFITHS
60 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
BRADLEY C. LOVE University College LondonXIAOJIN ZHU University of WisconsinKAUSTUBH PATIL University College LondonLUKASZ KOPEC University College London
This talk explores the interconnections between capacity constraints inmemory retrieval, the benefits of idealising training examples, and optimalteaching procedures. Optimal teaching procedures provide both anormative basis for idealisation and a means to evaluate competingcognitive models. The goal of this research is to understand why humandecision making goes wrong and to suggest methods to improveperformance.
Some decisions, such as predicting the winner of a baseball game, arechallenging in part because outcomes are probabilistic. When making suchdecisions, one view is that humans stochastically and selectively retrieve asmall set of relevant memories that provides evidence for competingoptions. We show that optimal performance at test is impossible whenretrieving information in this fashion, no matter how extensive training is,because limited retrieval introduces noise into the decision process thatcannot be overcome. One implication is that people should be moreaccurate in predicting future events when trained on idealised rather thanon the actual distributions of items. In other words, we predict the best wayto convey information to people is to present it in a distorted, idealisedform. Idealisation of training distributions is predicted to reduce the harmfulnoise induced by immutable bottlenecks in people’s memory retrievalprocesses. In contrast, machine learning systems that selectively weight (i.e.,retrieve) all training examples at test should not benefit from idealisation.These conjectures are strongly supported by several studies and supportinganalyses. Unlike machine systems, people’s test performance on a targetdistribution is higher when they are trained on an idealised version of thedistribution rather than on the actual target distribution. Optimal machineclassifiers modified to selectively and stochastically sample from memorymatch the pattern of human performance.
Optimal teaching procedures provide a means to test and extend thisbasic theory of capacity constraints and the benefits of idealisation. Givena test environment and agent, the optimal teacher determines the set oftraining examples that should maximise future performance. Critically, theoptimal set of training examples depends on the model of the agent. Wefind that limitedcapacity agents yield sets of training examples that areidealised, which provides a normative basis for idealisation manipulationsgiven the conclusion that humans are capacity limited. However, thesetraining sets have some characteristics that diverge in informative ways fromour initial intuitions about how to best idealise training examples to benefithuman decision makers. Finally, optimal teaching provides a novel methodto test competing cognitive models models should be favoured to theextent that they yield optimal training sets that maximise humanperformance. For example, we predict that optimal training sets derived fromlimitedcapacity models should yield better human performance than thosederived from unlimitedcapacity models. Overall, the reported resultssuggest firm limits on human rationality and have broad implications for howto train humans tasked with important classification decisions, such asradiologists (one test domain is mammography), baggage screeners,intelligence analysts, and gamblers (one test domain is sport prediction).
Keywords: Optimal Teaching, Memory Retrieval, Model Selection, Idealisation
Day 3 (Thursday) 10h30
DMB 2014 | 912 Sept 2014 | 61
Franc i s Cr ick Room
SHENGHUA LUAN Max Planck Institute for Human DevelopmentLAEL J. SCHOOLER Max Planck Institute for Human DevelopmentGERD GIGERENZER Max Planck Institute for Human Development
In a lexicographic semiorders model for preference, a decisionmaker isassumed to search cues in a subjective order and choose an alternative ifits value on a cue exceeds those of other alternatives by a threshold Δ. Wegeneralized this model from preference to inference and refer to it as Δinference. Unlike with preference, which is a matter of taste and for whichaccuracy is difficult to define, the problem a mind faces when making aninference is to select a Δ that can lead to accurate judgments.
To find a solution to this problem, we applied Clyde Coombs ’s theory ofsinglepeaked preference functions. We show that the accuracy of Δinference can be understood as an approachavoidance conflictbetween the decreasing usefulness of the first cue and the increasingusefulness of subsequent cues as Δ becomes larger, resulting in a singlepeaked function between accuracy and Δ. We refer to the peak of thisfunction as PeakΔ and found it varied with the properties of the taskenvironment: the more redundant the cues and the larger the differences intheir information quality, the smaller the PeakΔ.
Learning the PeakΔ in each environment can be costly. An alternativeway is to find a fixed Δ that works the best across a variety of environmentsand apply that Δ in a new environment. To gain an understanding of whatthat Δ might be, we started an investigation that involved 746 simulatedthreecue environments. In each environment, we examined both the fittingand prediction accuracy of Δinference under a series of fixed Δ values inthree samplesize conditions (N=20, 100, and 2,000). We found that forprediction and averaged across all simulated environments, the best fixedΔ is not large (i.e., 0.5 zscore) and it is the same for each samplesizecondition.
To check the generalizability of the finding, we collected 39 realworldenvironments and replicated the analysis there. To our surprise, the bestfixed Δ turned out to be zero across those environments! Using Δinferencewith Δ=0 means that a decision is made as soon as there is any differencebetween the alternatives on the first cue. In terms of prediction accuracy,this extremely simple model not only outperformed Δinference with anyother fixed Δ values, but also on average performed as well as Δinferencewith PeakΔ for each environment. We further compared Δinference with Δ=0with multiple linear regression (ordinary and Bayesian) and the generalmonotone model (Dougherty & Thomas, 2012), and found that it performedeither better than or at similar levels as the other models when the trainingset was less than 80% of the entire sample.
Overall, our study demonstrates the potential of integrating andextending existing concepts, models, and theories from preference toimprove our understanding of how the mind makes inferences. It also addsto the growing literature that shows how and why the mind can make fast aswell as accurate decisions using simple models that do not aim to optimize.
Keywords: Threshold, Δinference, lexicographic semiorders, singlepeakedfunction, approach–avoidance conflict, ecological rationality
Day 3 (Thursday) 10h30
62 | DMB 2014 | 912 Sept 2014
James Watson Room
CHRISTOPHE TANDONNET Université de Genève, Genève, Switzerland
Sequentialsampling models of decisionmaking differ according to twomajor distinctions: (i) how the evidence in favour of the different alternativesis accumulated, and (ii) what reflects the accumulation. First, accumulationof evidence may be independent for the different alternatives or interact,possibly through inhibition of the nonselected alternative. Second, whenconsidering decisions between actions based on sensory information,accumulation of evidence may reflect either a purely perceptual decisionthat leads thereafter to an action or a decision involving the selection ofan action. In the former case, the perceptual code has to be transmittedinto a motor code with either a sequential or continuous transmission.However, a fully continuous transmission can be viewed as functionallyequivalent as considering that decision involves action selection. Thus, atleast for this type of models, response implementation directly reflects theaccumulation of evidence, and hence can be used to assess the principlesof decision making.
In previous studies, the state of the motor system was assessed duringthe reaction time of a visual choice task involving manual responses. Eventrelated potential data are consistent with an inhibition of the nonselectedresponse: responserelated activity showed a positive component symmetricto the negative component involved in the motor command of the responsethat develops over the sensorimotor cortices. Transcranial magneticstimulation data showed that the motor activity related to the nonselectedresponse is reduced after an initial facilitation for both responses. Takentogether, these data are consistent with an initial activation of bothresponses followed by an inhibition of the nonselected response.
Such a pattern of response implementation provide constraints to themodels. Indeed, these data are consistent without further assumptions onlywith the racelike models (i) either assuming continuous transmission fromdecision to motor implementation or assuming that decision involves actionselection, and (ii) assuming inhibition between the different alternatives. Incontrast, the data are not compatible with the racelike models involvingindependent accumulation. The data constrain also the dualprocessmodels assuming sequential transmission from decision to motorimplementation. Although, by definition, response implementation cannotconstrain the decision in this case, at least two further assumptions arerequired to account for the activation and inhibition pattern: a nonspecificactivation independent of the decision process, and a specificfeedforward motor inhibition between decision and motor implementation.
In conclusion, measures of response implementation can provideconstraints to the models. One may speculate that this putative selectiveinhibition reflects an intrinsic principle of decision making. However, thisselective inhibition may not be generalized to the oculomotor system,although other form of inhibition like nonselective lateral inhibition may playa key role in generation of saccadic eye movements. Future researchassessing whether this selective inhibition can be evidenced in decisionsinvolving ocular responses would provide more constraints to dissociatebetween racelike models with inhibition, assuming either that decisioninvolves action selection or assuming continuous transmission from decisionto motor implementation, and models assuming sequential transmission.
Keywords: sequentialsampling models; vision; action; neural inhibition;sensorimotor; oculomotor
Day 3 (Thursday) 10h30
DMB 2014 | 912 Sept 2014 | 63
Rosa l i nd Frank l i n Room
ADNANE EZZIZI University of BristolDAVID LESLIE University of BristolSIMON FARRELL University of Bristol
Working memory (WM) tasks can be considered as dynamic decisionmaking tasks, where the agent needs to decide which action to take foreach new observation. The decision maker not only needs to use thecurrent observation, but also should take into account a sequence of pastobservations. The environment created by these types of tasks can be seenas partially observable Markov decision process (POMDP). This means thatit does have the Markov property—rewards depend only on the currentstate—but assumes that the underlying state of the environment cannot befully observed.
One approach to solve POMDPs is to support the learner with amemory device to store past states and/or actions in order todisambiguate the underlying state (McCallum, 1995). This method fits nicelywith WM tasks, as the memory device can simply be the agent ’s WM. Basedon this method, several reinforcement learning models (RL) of WM has beenproposed and shown to solve the learning problem of many WM tasks suchas the nback and the 12AX task (O'Reilly & Frank, 2006; Todd et al.,2009).
However, learning in these models is very slow, especially when the taskrequires more than one WM state, raising questions about whether WMcould reasonably be learned purely from experience. Using simulations withdifferent WM capacities, on a simplified version of the 12AX task, I will showthat there exists a tradeoff between learning speed and learningperformance, which might explain why our WM capacity is limited. I will alsopropose an approach for how an organism could deal with this tradeoff.Finally, I will propose several ways to increase the performance andlearning speed of these models. Hence, we can start testing and comparinglearning performance given by these models with human data.
Keywords: Working memory tasks; working memory capacity; Reinforcementlearning
Day 3 (Thursday) 11h00
64 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
KONSTANTINOS TSETSOS University of Oxford
Years of research within the cognitive and decision sciences haveconverged to the idea that humans are systematically at odds with rationalchoice theory. What is the cost of being irrational? Rationality is anecessary condition for optimality, the ability of an agent to choosecorrectly the course of action that maximizes the expected reward. Althoughirrational agents fail by definition to achieve the ideal observer benchmark,this exact performance difference cannot be experimentally measured dueto the subjective nature of “correctness” in preferential decisions. Tocircumvent this problem and inspired by research in perception and visualpsychophysics, I will present a novel paradigm (termed “valuepsychophysics”) that abstracts valuebased decisions into a simpleinformationintegration task in which correctness is objectively defined.Apart from offering precise control on the information preceding eachdecision, this task permits the orthogonal assessment of choice consistency(rationality) and choice accuracy (optimality). I will show that humanbehaviour violated the rational principles of transitivity and regularity but,surprisingly, the degree of these violations was not negatively correlatedwith choice accuracy.
From a theoretical perspective, choice accuracy is limited by twofactors a) reliance on nonnormative choice mechanisms and b)uncorrelated noise that distorts processing, offering an upper boundary onchoice accuracy. Why did the more irrational participants did not alsohave lower accuracy? This empirical result suggests that noise is a muchstronger predictor of poor performance compared to reliance on irrationaldecision mechanisms. To theorise this, I will describe a new sequentialsampling model that explains rationality violations in decisions underuncertainty using a single selective mechanism, which prioritises theintegration of the momentarily most valuable samples of incominginformation. Surprisingly, under this computational scheme, a higher tendencytowards irrationality (realised as stronger selectivity in the model) increasesrobustness to uncorrelated perturbations of the variable that drives thedecision (e.g. integrated evidence) or in other words to “ late” noise(corresponding for example to cortical noise). As a result, for moderatelevels of late noise the irrational selective integration model outperforms thechoice accuracy of the statistically optimal sequential probability ratio test(SPRT). I will present data from new experiments that aim to test thehypothesis that stronger reliance on irrational mechanisms helps (and doesnot deter) choice performance in the face of increasing late noise. Thisinterpretation of nonnormative behaviour downplays the severity of actingirrationally and offers a new explanatory framework of why humans evolvedirrational mechanisms.
Keywords:rationality, optimality, noise
Day 3 (Thursday) 11h00
DMB 2014 | 912 Sept 2014 | 65
James Watson Room
VALERIO BISCIONE Plymouth UniversityCHRISTOPHER HARRIS Plymouth University
Background. It is widely accepted that human reaction times (RT) canreveal important information about decision strategies for “allornone” typeof responses. During the past decades several models have beenproposed to describe and predict mean RT and RT distributions and,amongst them, sequential sampling models are the most successful inaccounting for data for simple and twochoice tasks. RT distributions areusually positively skewed which are often modelled as an Inverse Gaussian(the first passage time of a stochastic Wiener process) or an ExGaussian.However it has been found that latency distribution of saccades are veryclose to the Reciprocal Normal, meaning that they are normally distributedin the rate domain. This represents a challenge for most stochastic risetothreshold models, since these models can only achieve normality in the ratedomain with implausible parameters. We have already observed, in aprevious study, the normality of the distribution in the rate domain. In thatexperiment we used a twoforced choice paradigm with an easy/difficultcondition and an accurate/urgent instruction sets. We now investigate thepossibility to find similar results with a simple RT paradigm, exploring therelation between the foreperiod (FP) time and the Piéron’s Law in the ratedomain.
Methods. In this experiment (12 subjects, 3 blocks of 250 trials) wevaried the FP time and the luminance of the stimuli. The participants wereasked to press a button as soon as they saw the stimulus. We used 3 FPconditions (0.6, 1, and 2.4 seconds) and 5 luminance levels (0.42, 0.71,1.21, 2.06, 3.50 cd/m2).
Results. As expected, the relationship between RT and luminancefollowed Piéron’s Law, and the mean RT increased with mean FP. The ratedistributions approached normality with the increasing of the FP, which mayindicate that with shorter FP a second process was contaminating the data.We fitted different distributions to the data and we found that theReciprocal Normal provided a good fit. We investigated how the twoindependent parameters of the Normal Distribution varied across theFP/luminance conditions.
Discussion. We discuss several problems of the risetothreshold modelsand how a Reciprocal Normal Distribution is not consistent with somestochastic rise to threshold models. We propose a simple optimality modelin which reward is maximized to yield to an optimal rate and therefore anoptimal time to respond, exploring the connection between this model andthe speedaccuracy tradeoff. We also show how Piéron’s Law naturallyarises from this model. Our key claim is that the main goal of the humandecision process in simple decision tasks is to maximize the rate of reward.
Keywords: Reaction times, latency, Reciprocal Normal, Rate Domain, Piéron's Law,Foreperiod
Day 3 (Thursday) 11h00
66 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
MARCUS VINÍCIUS CHRYSÓSTOMO BALDO University of São Paulo, BrazilCAROLINA FEHER DA SILVA University of São Paulo, BrazilCAMILA GOMES VICTORINO University of Surrey, UK.EDUARDO SANGIORGIO DOBAY University of São Paulo, BrazilNESTOR CATICHA University of São Paulo, Brazil,
In repeated binary choice experiments, when choosing between twopossible alternatives that differ in their probability of occurrence, agentsbelonging to several different species, including humans, usually match theirchoice frequencies to the corresponding outcome probabilities, abehaviour known as probability matching. This strategy is suboptimal incomparison to always choosing the outcome with the higher probability(known as maximizing). We have previously shown that probability matchingmay be a likely byproduct of adaptive cognitive strategies that are crucialin sequence learning, but may lead to suboptimal performances in lessstructured environments (FeherdaSilva & Baldo, PLoS ONE 7(5): e34371).Here, we report on a simple but powerful analysis based on the computingof the conditional probabilities arising from the coupling betweenpredicted and actual binary sequences. By means of this method we wereable to show that, despite the common observed matching between thealternatives' probabilities and their respective choice frequencies, thedecisional behaviour of human agents clearly deviates from what would beexpected by their simply establishing probabilities of choice that match theunderlying probability structure of the two alternatives (as learned from aprevious sample of outcomes). In this sense, probability matching is a labelthat should better describe the empirical fact of roughly coincidentfrequencies, rather than an underlying decisional behaviour. Moreover, byapplying the same analytical method to human participants belonging todifferent age groups (from 3.9 to 71.2yearold volunteers), we were ableto characterize the explorationexploitation tradeoff across a widesegment of the human life span. Our results suggest that ageing tends toshift the decisional behaviour from a predominantly explorative procedure,as observed in children, towards a more exploitative strategy, which wasadopted by the elderly volunteers.
Keywords: decision making, probability matching, explorationexploitation tradeoff
Day 3 (Thursday) 11h30
DMB 2014 | 912 Sept 2014 | 67
Franc i s Cr ick Room
PERKE JACOBS Max Planck Institute for Human Development
There is an abundance of situations requiring decision makers tochoose from sequences of options rather than option arrays. This studyexamines the performance of three classes of satisficing heuristics identifiedearlier by Seale and Rapoport (1997). Structurally, all three classes ofheuristics are instances of the more general class of satisficing heuristics(Simon, 1955): Using different criteria, they formulate aspiration levelsbased on the first realizations of each sequence and then select the firstoption meeting the respective aspiration level. Seale and Rapoport (1997)demonstrated that experimental subjects ’ choices were consistent with allthree classes of heuristics.
Later, Stein, Seale and Rapoport (2003) evaluated the decision qualityof the three classes of heuristics in the context of the secretary problem, i.e.,a fully random sequence with no option to recall previously encounteredoptions. Given the fact that reallife sequences may not be fully random,the present study examines their performance A) when sequences areserially dependent; and B) when search costs are considered. Specifically,sequences were modeled as a multinomial process with fixed probabilitiesthat the sequence does not change, changes upwards, or changesdownwards by a fixed amount. The parametrization of the sequence wasbased on price data from a German online market for used automobiles(Artinger, 2012). This simulation study followed a full factorial design inwhich probabilities of price increase and decrease, the magnitude of pricechanges, and search costs were varied. Gilbert and Mosteller (1966)provided an optimal solution for secretaryproblemtype of situations, whichwas used as a performance benchmark by Stein and colleagues. Incontrast, the present study employed a competitive testing strategy:Performance of the heuristics was assessed relative to the performance of aBayesian decision algorithm that is aware of the statistical structure of thesequence and inferred parameter values from observed realizations of thesequence.
The results indicated that the single best heuristic from each classdiffered depending on the trend of the sequence, with improvingsequences requiring longer and deteriorating sequences requiring shortersampling periods. Given the best performing heuristic for each class,performance was compared to that of the Bayesian algorithm acrossseveral conditions. The Bayesian algorithm offered the best performanceacross several conditions — perhaps not surprisingly, given its relativesophistication. Interestingly, all classes of satisficing heuristics offeredconsiderably better performance than the Bayesian algorithm when 1)sequences exhibited an improving trend; and 2) search costs were zero sothat performance differences stemmed only from finding the best option.Decision environments with these characteristics therefore allow decisionmakers to employ a simple, psychologically plausible strategy whoseperformance exceeds that of a computationally more intensive Bayesianalgorithm. These results reaffirm Simon's (1956) conjecture that heuristicscan be effective decision strategies when adapted to the specificdecision environment.
Keywords: Satisficing, Heuristics, Sequential Choice, Competitive Testing
Day 3 (Thursday) 11h30
68 | DMB 2014 | 912 Sept 2014
James Watson Room
ANGELO PIRRONE University of SheffieldJAMES A.R. MARSHALL University of SheffieldTOM STAFFORD University of Sheffield
The DDM has been shown to be quantitatively accurate in describingbinary choices and valuebased choices. According to the DDM thedecisionmaker integrates the difference in evidence supporting twoalternatives. In laboratory settings subjects are usually rewarded only onmaking a correct choice, so optimisation of a zeroone loss function isappropriate, and this is achieved by implementing a statisticallyoptimaldecision procedure that gives the best compromise between speed andaccuracy of decisionmaking; this tradeoff can be implemented andoptimised by the DDM. However many decisions, such as selecting fooditems of potentially different value, appear to be different since the animalis rewarded by the value of the item it chooses regardless of whether it wasthe best. We argue that most naturalistic decisions, which animals ’ brainsshould have evolved to optimise, are valuebased rather than accuracybased. While valuebased decisions can be optimised using mechanismsfor managing speedaccuracy tradeoffs, this requires additionalinformation on the decision problem at hand in order to allow optimalparameterisation since subjects should learn the values of correct andincorrect choices over time on a case by case basis. Recent theory by Paiset al. (2013) has presented mechanisms that can manage valuesensitivedecision problems adaptively without direct parameterisation regardingeach alternatives ’ values. In Pais et al.'s model when equalbutlowvaluealternatives are presented, a decision deadlock is maintained that can bebroken should a third, highervalue alternative, be made available whilewhen equalbuthighvalue alternatives are presented, or sufficient timepasses, deadlock is spontaneously and randomly broken; when differencesbetween alternative values are sufficiently large, the valuesensitivemechanism becomes closer to a classic DDM, allowing speedaccuracytradeoffs to be managed. In contrast, the DDM is insensitive to theabsolute magnitude of evidence for alternatives, with its behaviourdetermined solely by their difference and by sensory noise. To discriminatethese two decisionmaking models we conducted two studies, one aperceptualjudgement accuracybased decision and one a valuebaseddecision. The perceptual experiment involved a numerosity discriminationparadigm, with subjects instructed that only correct identifying the largestdotcluster would be rewarded; in the valuebased setting we adjusted theparadigm by instructing participants that they would receive a monetaryreward in direct proportion to the sum of the dots in the clusters theyselected over all trials. The main results of the two studies are (i) in line withthe DDM, reaction times differ according to the difference in evidence ofthe two alternatives, (ii) in contrast with the DDM and in line with Pais ’ modelpredictions, reaction times differ according to the overall magnitude of thetwo alternatives for each level of ‘difference in evidence’. Our theoreticalconsiderations and the initial results from the study, that cannot be fitted bythe DDM, raise issues with the idea that the DDM can be a comprehensivecomputational framework for decision making and suggest that humandecision making is indeed value sensitive, even in simple perceptual tasks.
Keywords: decisionmaking, value, reward, driftdiffusion, mechanism, evolution,optimality
Day 3 (Thursday) 11h30
DMB 2014 | 912 Sept 2014 | 69
Rosa l i nd Frank l i n Room
JOHN GROGAN University of BristolRAFAL BOGACZ University of OxfordELIZABETH COULTHARD University of Bristol; North Bristol Trust NHS
Dopamine has been found to improve learning from positivereinforcement at the cost of negative reinforcement (e.g. Frank et al., 2004),but has also been suggested to work mainly through effects on testing andretrieval rather than learning (e.g. Shiner et al., 2012). We aimed toestablish whether the dopamine levels affect encoding or retrieval ofstimulusaction associations learnt from positive (or negative) feedback.Our paradigm included a 24 hour delay between learning and testing, thatallowed Parkinson’s Disease (PD) patients to be tested on or off theirdopaminergic medication on both days (giving 4 conditions: onon, onoff,offon, offoff). This within subjects design was tested on 15 PD patients and13 healthy agematched controls. On day 1 participants were givenlearning trials of the probabilistic selection task in which the feedbackgenerated after selecting one of two cards is probabilistic. For example,card A receives positive feedback on 80% trials while card B receives it on20% (65% & 35% for C & D, respectively). After learning, memory for thesecards is tested (without feedback given) both immediately and after a 30minute delay. On day 2 participants were given another memory test of thecards, and then shown the new combinations they hadn’t seen previously(AC, AD, BC, BD). The amount of times they chose A (the most rewardedcard) or avoided B (the most punished card) reflected learning frompositive and negative reinforcement, respectively. We also fit a variety ofreinforcement learning models to the patient data, both to see whichmodel best explained the data and to see any effects of dopamine on thelearning rates of the patients. We found a significant interaction of day 1and day 2 medication states on the amount of chooseA behaviour. Whenpatients were off medication on day 1, day 2 medication increasedchooseA and decreased chooseB as expected (offon), relative to day2 off medication (offoff). However, when patients were on medication onday 1, day 2 medication had no effect. Intriguingly, patients on medicationfor both days (onon) showed the highest amount of avoidB and lowestamount of chooseA behaviour of any of the conditions, suggesting thatthe effects of dopamine are very different when tested 24 hours latercompared to immediately or only 1 hour after learning. The modellingrevealed that the best fitting model (as determined by the BIC) was a duallearning rate SARSA model, but there were no significant effects ofmedication on the learning rates. These findings suggest that dopamine hasa different pattern of effects when tested 24 hours after learning, and thatdopamine state during learning determines whether later dopamine willhave an effect.
Keywords: dopamine, reinforcement learning, parkinson's disease
Day 3 (Thursday) 12h00
70 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
PAULA PARPART University College LondonMATT JONES University College LondonBRADLEY C. LOVE University College London
In the judgement and decision making literature, rational inferencemodels (e.g. Bayesian models) have usually been depicted as opposingsimple heuristics (e.g., takethebest or tallying). The current researchdemonstrates a novel finding wherein prominent decision heuristics are aspecial case of Bayesian inference.
We developed two Bayesian learning models based on two wellknownregularization techniques in machine learning, namely lasso regression andridge regression. The penalty terms of these Bayesian models eachincorporate a prior that reflects the necessary environmental structures fortallying or takethebest to succeed. For example, tallying performs best inan environment where cues are equally weighted, which is sometimesreferred to as a compensatory weighting structure. In contrast, takethebestthrives when the most heavily weighted cue outweighs all other cues suchthat no combination of theirs can compensate for the strongest weight,often referred to as a noncompensatory structure. We demonstrate thatthese Bayesian inference models become equivalent to the heuristics whenthe priors become very extreme (with an infinitely large penalty parameter),thereby making them a special case of the model, next to standard linearregression.
In a reanalysis of popular heuristic datasets which span a wide rangeof domains we show that our modified Bayesian ridge regressionoutperforms both tallying and standard linear regression. Similarly, ourBayesian extension of lasso regression could outperform both takethebestand linear regression. A large simulation study furthermore illustrates thepractical convergence of the Bayesian learning models and the heuristicswith increasingly extreme priors: convergence is reached when the originalcue weights have developed into perfectly compensatory (tallying) ornoncompenstary (takethebest) weighting structures. Results depict that theBayesian learning models declare both heuristics and standard linearregression as a special case of the Bayesian inference model. This impliesthat heuristics can be adaptive to certain environments, while beingcontinuously contained within a rational inference model. Thereby wecreate a formal relationship between two traditionally opposing theories ofdecision making.
In addition, these findings lead us to think that sometimes theappropriate psychological process and its matching environmentalstructure may lie somewhere in between a frugal heuristic and more complex,integrative regression approach. These new developments have farreaching implications with respect to the judgement and decision makingliterature on heuristics and rational inference models, as well as other fieldsthat make use of novel regularization algorithms.
Keywords: heuristics, Bayesian inference, takethebest, regularized regressions,ridge regression
Day 3 (Thursday) 12h00
DMB 2014 | 912 Sept 2014 | 71
James Watson Room
TAKAO NOGUCHI University of WarwickBRADLEY C. LOVE University College London
Sequential sampling models provide a way to understand both speedand accuracy of judgement. In these models, an individual is assumed toaccumulate evidence supporting a judgement, and when the accumulatedevidence reaches a response criterion, the individual makes a judgement.For example, in exemplarbased randomwalk (EBRW) model (Nosofsky &Palmeri, 1997), items are retrieved from memory sequentially and theircategory membership provides evidence for the respective response. Theprobability of retrieving an item depends on its similarity to the stimulus.
Accumulation rate in sequential sampling models is typically stationaryover time. In EBRW model in particular, the probability of recalling aparticular item is fixed until an accumulation reaches a response criterion.As an individual learns more about structure of a stimulus, however, the waythis individual processes information may be altered over time. Previousstudies report that especially when under time pressure, an individual is likelyto evaluate an alternative as attending attribute dimensions (e.g., Lamberts,1995). These findings indicate that accumulation rate may be based on asubset of dimensions at first, and gradually, more dimensions becomeincorporated. In other words, the similarity of a stimulus and items in memorymay change within a trial.
To implement the process where a dimension is evaluated as attended,we propose a dynamic construction of similarity perception. Here, anindividual recalls an item similar to the stimulus at hand based only on thedimensions the individual has attended. On the first fixation, similarity is afunction of only the one dimension, but with time, similarity becomes basedon more dimensions. This dynamic construction predicts that when anindividual first attends to a stimulus dimension that provides misleadinginformation about category membership, this individual accumulatesevidence for an incorrect judgement at first. As a result, 1) the individualshould be slower to make a correct judgement, as the individual has tooverride initially accumulated evidence, and also 2) the individual shouldbe less likely to make a correct judgement.
These predictions concerning the order in which dimensions aresampled were empirically tested with eyemovement recordings. In theexperiments, participant learned to classify an amoebalike organism withthree dimensions. In classifying these stimuli, all three dimensions arerelevant but are imperfect predictors (Type IV structure; Shepard, Hovland,& Jenkins, 1961). Even when a stimulus belongs to Category A, for example,its shape at the top may be more readily shared with the stimuli in CategoryB. When such unrepresentative dimension is fixated at first, compared towhen the same dimension is fixated later, participant takes longer tocorrectly classify the stimulus, and is less likely to make a correctclassification judgement.
Thus, our results support the dynamic construction of similarityperception. We computationally formulate this construction and proposean extension to the EBRW model. This proposed model is comparedagainst existing models in their predictive accuracy. We conclude with adiscussion on implementation of dynamic processes to other sequentialsampling models.
Keywords: Sequential sampling; process tracing; exemplar
Day 3 (Thursday) 12h00
72 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
TIMOTHY MULLETT University of WarwickRICHARD TUNNEY University of NottinghamNEIL STEWART University of Warwick
Attention is an important, yet often overlooked topic in decision makingresearch. Many models make specific predictions about which items andattributes an individual is likely to attend to and at what point in thedecision process. There are also other models which are not intentionallyformulated to predict attention, but still imply that specific information will bedeliberately ignored or that the decision process will terminate after aspecific piece of information is made available to the model ’s deliberativecore. We use eyetracking to investigate a prediction which applies toseveral models: that individuals spend more time attending to informationwhich they weight more strongly in their choices. Furthermore, we examine theprediction that this bias will increase over time as individuals move from aninitial exploratory, information encoding phase to a second, moredeliberative strategy. Previous research has demonstrated that in choicesbetween simple, single attribute items, individuals have a bias to attendmore to the item they eventually choose and that this bias grows over time.This is a phenomena referred to as the “gaze cascade”. Our study is the firstto examine these effects of attention bias within multiattribute choice.
The results show that the gaze cascade effect is remarkably robust. Theeffect size is not significantly different to previous demonstrations, despitethe significant additional complexity of items with 5attributes each.However, the results also show that there are no attributewise effects:Individuals do not attend more to the information that they subsequentlyweight more highly in their choices. In fact we find no attributewise attentionbias, with attention proportions not differing significantly from chance evenwhen their influence on individuals ’ choices differs by an order of magnitude.
We also present a series of simulations which demonstrate that theexistence of the gaze cascade is only compatible with models which positthat evidence is accumulated over time and that a decision is only madewhen the relative difference in accumulated evidence between the bestand second best items becomes sufficiently large. The results are robust tovery significant changes in the core model, so long as these twoassumptions are retained. The findings from our computational simulationsand empirical research provide significant constraints for existing and futuremodels.
Keywords: Attention, value, choice, eyetracking, modelling
Day 3 (Thursday) 14h00
DMB 2014 | 912 Sept 2014 | 73
Franc i s Cr ick Room
ELLIOT A. LUDVIG University of WarwickCHRISTOPHER R. MADAN University of AlbertaMARCIA L. SPETCH University of Alberta
When making decisions from experience, people often exhibit differentpatterns of risky choices than when making analogous decisions fromdescription. For example, when asked whether they would prefer aguaranteed $20 or a 50/50 chance of winning $40, people are typicallyrisk averse. In contrast, when the same risks and rewards are learned fromexperiences, people are often more risk seeking.
Here, we present data from a series of experiments that identify threefactors that further increase risk seeking in decisions from experience: bigwins, fast responses, and reminders for past wins. These factors may provekey to understanding why people engage in realworld gambling.
In Exp 1, people were tested on a pair of decisions. On highvalue trials,they chose between a fixed 60 points and a 50/50 chance of 40 or 80points. On lowvalue trials, they chose between a fixed 20 points and a50/50 chance of 0 or 40 points. People were more risk seeking on the highvalue trials than the lowvalue trials, reflecting an extra sensitivity to thebiggest win (80) and the smallest win (0).
Exp 2 used an identical protocol as Exp 1, except there were twoseparate groups: fast and slow. People in the fast group had a shortdeadline for responding (1.5 s) and short intertrial intervals (0.5 s). The slowgroup had a longer deadline (5 s) and longer intertrial intervals (4 s). Forboth low and highvalue decisions, people were more risk seeking in thefast group, when they had to make their decisions rapidly.
In Exp 3, people were repeatedly tested on a decision between a fixed40 points, and a 50/50 chance at 20 or 60 points. Each of the outcomeswas also directly associated with a unique image. Before some of thedecisions, people were primed with the unique image, thereby remindingthem of the past outcome. Reminders for past wins increased risk seeking bynearly 20%, while reminders for the other two outcomes did not alter riskychoice.
We interpret these results as reflecting a samplingbased decisionprocess, along the lines of the decisionbysampling (DbS) framework or theDyna algorithm from reinforcement learning. On this view, people makedecisions by sampling from the history of past outcomes and comparing thesampled outcomes for each option. These samples can be biased towardparticular outcomes, either implicitly (e.g., saliency of big wins in memory) orexplicitly (e.g., through reminders), allowing people’s risky choice to shiftbased on the context.
Keywords: Risky choice, decisions from experience, gambling, extreme outcomes,time pressure, memory biases
Day 3 (Thursday) 14h00
74 | DMB 2014 | 912 Sept 2014
James Watson Room
CLAIRE HALES University of BristolEMMA ROBINSON University of BristolCONOR HOUGHTON University of Bristol
Cognitive biases influence many aspects of human decision making.Some biases have also been shown in animals, such as judgement biases ininterpretation of ambiguous information. This bias can also be altered byaffective state. However, the cognitive processes underlying this cognitiveaffective bias (CAB) have not been explored. In rodents, a twochoicereaction time ambiguous cue interpretation task has been used to studyCAB. Modelling data from twochoice reaction time tasks using the diffusionmodel can provide insight into cognitive aspects of the decision process,each of which is represented by a different parameter. We aimed toempirically validate use of the diffusion model with this ambiguous cueinterpretation task by using specific manipulations that should affectcorresponding model parameters of interest.
Male lister hooded rats (n=14) were trained to discriminate betweentwo distinct auditory tones and make a response on the appropriate leverto either obtain high value (four sugar pellets) or low value (one sugarpellet) reward. Following successful discrimination, CAB was measured in testsessions where ambiguous tones and reference tones were presented.Tones with varying ambiguity were used to test the capacity of the diffusionmodel to represent differences in stimulus discriminability, which should affectthe drift rate parameter. To test if the diffusion model is sensitive to changesin starting bias, reward value was altered so that a correct lever press aftereither tone resulted in a single sugar pellet reward. This should eliminate anydifferences in response bias between the two tones.
Experimental manipulations caused expected changes in correspondingdiffusion model parameters that matched with behavioural data. The driftrate for the midpoint ambiguous tone was closest to zero, indicating it wasthe most difficult to discriminate. Nearmidpoint ambiguous tones had driftrates that were not significantly different to the corresponding referencetone (p’s>0.05), matching behavioural data that indicated comparablediscrimination for these tones as for reference tones. Tones associated withthe high value reward had significantly positively biased starting pointparameters (p’s<0.001). Interestingly, despite apparent neutral respondingto the ambiguous midpoint tone from behavioural measures, the diffusionmodel indicated a significant positive bias in starting point for this tone(p=0.027). There were no significant differences between reference tonestarting point parameters (p=0.472), or in the magnitude of the drift rateparameters (p=0.948) when both tones predicted the same outcome.
This study provides initial validation that the diffusion model can beapplied to data from a rodent ambiguous cue interpretation task used tomeasure CAB. Experimental manipulations that should specifically affectstimulus discriminability and response bias did alter correspondingparameters in the diffusion model. Furthermore, application of the diffusionmodel to this CAB task can expose subtle changes in behaviour that aremasked through traditional data analyses of behavioural measures. Thissupports further use of the diffusion model with this task to investigate howthese aspects of the decision making process are altered by affective statemanipulations and relate to changes in cognitive bias.
Keywords: cognitive affective bias, diffusion model, rodent
Day 3 (Thursday) 14h00
DMB 2014 | 912 Sept 2014 | 75
Rosa l i nd Frank l i n Room
ROSIE CLARK University of BristolRAFAL BOGACZ University of OxfordIAIN D. GILCHRIST University of Bristol
Prior information about our environment influences our decisions aboutwhere to look, and how urgent those decisions are. The relative probabilityof stimuli and the reward associated with them are two highly influencingforms of prior knowledge. Saccadic reaction times are faster to locationsthat occur more frequently (Carpenter and Williams, 1995), and some(largely primate) studies have shown mean saccadic latency was shorter inrewarded trials than in nonrewarded (Takikawa et al 2002). However, thereis a lack of spatial reward contingencies in human studies (Bendiksby &Platt 2006). Within models of saccade latency distributions (such asLATER) both effects could be accounted for by a modulation in the startingpoint of the accumulation process.
In the first experiment we investigated the effect of spatial probabilityon manual response time and saccade latency concurrently. We found asignificant difference between the saccadic response times across threeprobability levels, but no effect in the manual response time.
In a second experiment we investigated the effect of reward in a similarparadigm, we manipulated both the spatial distribution of reward within ablock of trials (high reward and low reward side) and the overall rewardlevel within a block (high reward block and low reward block). We found asignificant effect of target side on the manual responses, and a significantinteraction between block type and target, suggesting that the context ofthe reward values has an effect. For saccade latency there was also asignificant interaction.
These results suggest that there are subtle differences in the way thatreward and spatial probability have an impact on saccadic and manualresponses respectively.
Keywords: saccades, reward, spatial probability, response times
Day 3 (Thursday) 14h30
76 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
LUKASZ WALASEK University of WarwickNEIL STEWART University of Warwick
One of the key assumptions in prospect theory is loss aversion. Lossaversion is the property of having a steeper value function for losses thangains. Surprisingly few attempts were made to elicit the loss aversionparameter λ in decision under risk. These estimates show great variability,with values of λ varying from 1.00 to 2.63. λ is the ratio of the slopes forlosses and gains, with a value of 1 indicating no loss aversion, and valuesgreater than one indicating loss aversion.
In four experiments we demonstrate that loss aversion results from theevaluation of a given gain or loss against the distribution of gains andlosses within the experiment. We focus on one parametric methods foreliciting loss aversion, which requires an individual to make a series ofaccept/reject decisions about 5050 gambles involving a gain and a lossof money (Tom et al., 2007). We experimentally manipulate the ranges forthe distributions of gains and losses in the experiment.
In decision by sampling theory (Stewart, Chater &, Brown, 2006) thesubjective value of a given gain/loss is given by the number of favourablecomparisons to gains/losses in memory. For example a gain of £10 will havea subjective value of 1/4 in an experiment with a uniform distribution ofgains in the range £0£40 because 1/4 of gains in this range are smallerthan £10 and make £10 look good. So decision by sampling predicts, inadvance, that we should see more loss aversion in an experiment with asmall range of losses and a large range of gains. For example, themagnitude £10 will have a subjective value of 1/4 when compared againstgains in the range 040 and a subjective value of 1/2 when comparedagainst losses in the range 020. This is loss aversionthe subjective valuefor a fixed magnitude is larger when it is a loss. If we were to reverse theranges of gains and losses, decision by sampling predicts that we shouldobserve the opposite of loss aversion.
In all four of the experimentsthree online and one lab experiment withtrue incentiveswe find that the loss aversion parameter is indeed largelydetermined by our choice of the distribution of gains and losses. Wereplicate loss aversion when the range of the gains is twice as large as therange of losses. When the ranges of gains and losses are equal we find noloss aversion. When the range of losses is twice as large as the range ofgains we find the opposite of loss aversion. We conclude that while theaversion to losses may be a robust empirical phenomenon, loss aversionresults, at least in part, from differences in the distributions of gains andlosses in the environment. We offer a novel explanation of the origin of lossaversion, proposing that loss aversion should not be interpreted as astable parameter of the prospect theory but instead as a property of theexperimental design.
Keywords: loss aversion, Decision by Sampling, relative judgments, context,decision under risk, prospect theory
Day 3 (Thursday) 14h30
DMB 2014 | 912 Sept 2014 | 77
James Watson Room
PETE TRIMMER University of BristolELIZABETH PAUL University of BristolMIKE MENDL University of BristolJOHN MCNAMARA University of BristolALASDAIR HOUSTON University of Bristol
Moods can be regarded as fluctuating dispositions to make positiveand negative evaluations. Developing an evolutionary approach to moodas an adaptive process, we consider the structure and function of suchstates in guiding behavioural decisions regarding the acquisition ofresources and the avoidance of harm in different circumstances. We use adrift diffusion model of decision making to consider the information requiredby individuals to optimise decisions between two alternatives, such aswhether to approach or withdraw from a stimulus that may be life enhancingor life threatening. We show that two dimensions of variation (expectationand preparedness) are sufficient for such optimal decisions to be made.These two dispositional dimensions enable individuals to maximize theoverall benefits of behavioural decisions by modulating both the choicemade (e.g., approach/withdraw) and decision speed. Such a structure iscompatible with circumplex models of subjectively experienced mood andcore affect, and provides hypotheses concerning the relationships thatoccur between valence and arousal components of mood in differingecological niches. The approach is therefore a useful step toward beingable to predict moods (and the effect of moods) using an optimalityapproach.
Keywords: Mood, emotion, evolution, core affect, circumplex, modelling.
Day 3 (Thursday) 14h30
78 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
LUIGI ACERBI University of EdinburghSETHU VIJAYAKUMAR University of EdinburghDANIEL M. WOLPERT University of Cambridge
Uncertainty is a ubiquitous feature in shaping human sensorimotorbehaviour (Orban & Wolpert, 2011). Notably, humans are able to integrateprior information with sensory data to build 'posterior' estimates ofperceptual and sensorimotor variables that take uncertainty into account(Kording & Wolpert, 2004).
In the first part of the talk, we show how combining probabilistic beliefsof varying complexity with a cost function leads to 'nearoptimal biases'modulated by the detailed form of the posterior, in a study of sensorimotortiming (Acerbi et al., 2012). Optimal performance, though, is not alwaysachievable when the task at hand requires the integration of complexprobabilistic information, and suboptimal biases may emerge. Crucially,such biases appear to be affected by estimation uncertainty too, asquantified by the spread of the observer's posterior distribution (Acerbi etal., 2012).
In the second part of the talk, we identify a possible source of suboptimal performance in the failure of the motor system to correct for errors inconditions of increased estimation uncertainty. To test our hypothesis,unbeknownst to participants we randomly shifted visual feedback of theirfinger position during reaching in a centre of mass estimation task. Eventhough they were given enough time to compensate for this perturbation atthe end of their movement (adjustment phase), participants only fullycorrected for the induced error on trials with low uncertainty about targetlocation; instead, correction was partial in conditions involving moreuncertainty.
Albeit suboptimal with respect to task demands, this lack of correctioncan be explained by considering an additional cost of adjusting one'sresponse in conditions of uncertainty, a term in the cost function that canbe interpreted as 'effort' (whether energy, time or computational). Even forsimple, naturalistic tasks, such as centre of mass estimation, the effect of thisadditional cost can be significant and is strongly modulated by trialuncertainty. Our findings suggest that subjects' decision uncertainty, asreflected in the width of the posterior, is a major factor in determining howtheir sensorimotor system responds to errors, supporting theoretical models inwhich the decision making and control processes are fully integrated.
Keywords:Sensorimotor learning, Bayesian Decision Theory, Human performance,Probabilistic inference, Suboptimality, Motor planning, Error correction.
Day 3 (Thursday) 15h00
DMB 2014 | 912 Sept 2014 | 79
Franc i s Cr ick Room
DAVID RONAYNE University of Warwick
I offer a theoretical account of how individuals evaluate multiattributealternatives along with empirical evidence. The model extends anddevelops the singleattribute decisionbysampling (DbS) model of Stewartet al. (2006). The multiattribute DbS model of how individuals assign valueto multiattribute alternatives assumes the following twostage process:1) The choice set presented to an individual is used as information toupdate the individual's belief over the total market offerings of thatalternative (or product); 2) an alternative within the choice set is thenevaluated by a finite series of dominance comparisons between it, theremaining alternatives in the choice set and others drawn from the posterior.The model captures all three of the most wellevidenced multiattributecontext effects: the attraction, compromise and similarity effects. Explainingall three of these effects within the same framework has been a challenge inthe decisionmaking literature. I show how each of these "big three" contexteffects are produced theoretically by the model. I then provideexperimental evidence, from both laboratory and online environments, thattests the model's predictions.
Keywords: Context Effects, DecisionbySampling, Amazon Mechanical Turk,Consumer Choice
Day 3 (Thursday) 15h00
80 | DMB 2014 | 912 Sept 2014
James Watson Room
CAROLINA DORAN University of BristolNIGEL R. FRANKS University of Bristol
Social insect colonies provide some of the richest examples of complexsystems in nature. They are an excellent model for experimental investigationinto questions of how group decisions are made as they allow directmanipulation of their components and observation of the collectivebehaviour. Temnothorax albipennis colonies are able to allocate theappropriate effort into gathering information regarding new homes inaccordance with the quality of the nest they currently inhabit. Furthermorewhen faced with a risky choice they seem to be risk takers and to ‘gamble’ ifthe expected payoff is positive, i.e. it represents a gain. However, the time ittakes colonies to reach a consensus gradually increases when the gain issmaller. Our latest research shows evidence that this species of ants exhibitseconomic rationality both at the individual and collective level. Bothindividual workers and colonies seem to base their decisions on the finalexpected payoff, namely the value of the target nest minus the value of thecurrent nest minus emigration costs. We so concluded that when they arenot totally isolated from other colony members and when they are not undersevere time pressure, even the so called – simple components: the individualworkers – are able to behave rationally. It is clear that animal collectivescan solve more complicated problems than their isolated individualmembers. Nevertheless, research on collective decisionmaking should notunderestimate the sophistication of the individual. We suggest that one ofthe clichés of complexity science – the stupid component and thesophisticated whole – needs to be overturned in certain cases, especiallyin biology, by a recognition of sophistication at all levels.
Keywords: risk, collective behaviour, decisionmaking
Day 3 (Thursday) 15h00
DMB 2014 | 912 Sept 2014 | 81
Rosa l i nd Frank l i n Room
DANIEL WOLPERT Royal Society Noreen Murray Research Professor &Professor of Engineering, University of Cambridge
The fields of decision making and sensorimotor control havedeveloped in parallel although both require acting in real time onstreams of noisy evidence. I will review our recent work coveringboth probabilistic models of decision making and sensorimotorcontrol and the interactions between these processes. This willinclude the development of a cognitive tomography techniquethat can extract complex priors from simple decision makingmeasurements. I will also review our work on the relation betweenvacillation and changes of mind in decision making, thebidirectional flow of information between elements of decisionformations (such as accumulated evidence) and motor processes(such as reflex gains) and how active sensing through eyemovements is used for visual texture categorisation.
Day 3 (Thursday) 16h45
DANIELWOLPERT
82 | DMB 2014 | 912 Sept 2014
DMB 2014 | 912 Sept 2014 | 83
Day 4 (Friday) 09h00Rosa l i nd Frank l i n Room
ANTONIO RANGEL Bing Professor of Neuroscience, Behavioral Biology andEconomics , California Institute of Technology
Neuroeconomics studies what are the computations made bythe brain in different decision situations, and how are thesecomputations implemented and constraints by the underlyingneurobiology. This talk describes recent fMRI, EEG and eyetracking experiments designed to understand how the braincomputes and compares values during simple decisions, likechoosing between an apple and an orange.
ANTONIORANGEL
84 | DMB 2014 | 912 Sept 2014
Day 4 (Friday) 10h15Rosa l i nd Frank l i n Room
DANIEL A. BRAUN Max Plank Institute for Biological Cybernetics and Max PlankInstitute for Intelligent SystemsJORDI GRAUMOYA Max Plank Institute for Biological Cybernetics and Max PlankInstitute for Intelligent SystemsPEDRO A. ORTEGA Max Plank Institute for Biological Cybernetics and Max PlankInstitute for Intelligent Systems
Both sensorimotor and economic behavior in humans can beunderstood as optimal decisionmaking under uncertainty specified byprobabilistic models. In many important everyday situations, however, suchmodels might not be available or be ambiguous due to lack of familiaritywith the environment. Deviations from optimal decisionmaking in the face ofambiguity have first been reported by Ellsberg in economic choicesbetween urns of known and unknown composition. Here we designed an urntask similar to Ellsberg's task and an equivalent motor task, where subjectschoose between hitting partially occluded targets with differing degree ofambiguity. In both experiments subjects had to choose between a risky andan ambiguous option in every trial. The risky option provided full informationabout the probabilities of the possible outcomes. The ambiguous optionwas always characterized by a lack of information with respect to theprobabilities. We could manipulate the degree of ambiguity by varying theamount of information revealed about the ambiguous option. In the motortask, we manipulated the extent to which an ambiguous target wasoccluded that subjects aimed to hit, whereas in the urn task we varied thenumber of balls drawn from the ambiguous urn before subjects made theirdecision. This way, we could test the more general hypothesis that decisionmakers gradually switch from ambiguity to risk when more informationbecomes available. Ellsberg's paradox then arises in the limit case in whichthe ambiguous option gives away no information. We found that subjectstended to avoid ambiguous urns in line with Ellsberg's results, however, thesame subjects tended to be ambiguityloving or neutral in the motor task.One of the most important points of Ellsberg's original experiment was toshow that expected utility models—that is models that only care aboutmaximizing expected success—cannot explain subjects' choice behaviorunder ambiguity. Since then a number of models for decisionmaking underambiguity have been proposed. However, few of them are able todynamically change the degree of ambiguity as new information arrives.Here we employ a multiplier preference model, that is a type of variationalpreference model for decisionmaking under ambiguity, and use it under aBayesian update procedure to integrate novel information. We show thatthe deviations from optimal decisionmaking can be explained by such arobust Bayesian decisionmaking model. Our results suggest that ambiguityis a ubiquitous phenomenon, not only to understand economic choicebehavior, but also sensorimotor learning and control.
Keywords: ambiguity, risk, sensorimotor learning
DMB 2014 | 912 Sept 2014 | 85
James Watson Room
RUSSELL GOLMAN Carnegie Mellon UniversityDAVID HAGMANN Carnegie Mellon UniversityJOHN H. MILLER Carnegie Mellon University
Decentralized systems of agents must often make critical choices, as, forexample, when a swarm of bees (or a colony of ants) chooses a new hive(nest) location. Such natural decentralized systems seem to share acommon decisionmaking process involving information search with positivefeedback and consensus choice through quorum sensing. We model thisprocess with an urn scheme that runs until hitting a threshold. We show thatthere is an inherent tradeoff between the speed and the accuracy of adecision. The proposed mechanism provides a robust and effective meansby which a decentralized system can make reasonably good, quickchoices. We find that this decision mechanism naturally produces riskaversion. Along with providing a deeper understanding of the behavior ofmany natural systems, this model provides a blueprint that could be appliedin the creation of novel and productive decentralized decision mechanismsfor use in human and artificial systems.
Keywords: consensus decision making, decentralized systems, Polya urn process,quorum, ants, bees
Day 4 (Friday) 10h15
86 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
A. ALDO FAISAL Imperial College LondonANASTASIA SYLAIDI Imperial College London
A rapidly growing body of evidence suggests that human decisionmaking in the context of motor behavior is consistent with OptimalFeedback Control (OFC) (Landy & Wolpert, 2012, Todorov & Jordan,2002, Nagengast et al, 2009). However the representations andmechanisms that underlie learning optimal controllers remain elusive,particularly in complex motor tasks that involve the manipulation of objects.We propose here that in such tasks the brain makes continuous decisionsfor the generation of complex trajectories by learning OFC based on theidentification of unknown parameters of both body and object dynamics(Sylaidi & Faisal, 2012). We present experiments and a novel theoreticalframework that can capture the temporal dynamics of motor learning on atrialbytrial basis.
Human subjects (N=15) were instructed to move a virtual object ofunknown dynamics from start to target locations in an unintuitive task thattranslated hand velocity into control forces on the object state. Subjectsreceived a performance feedback after the completion of each trial in theform of a cost function capturing the end point error, a control penalty andan end state penalty.
While learning an optimal controller directly is a sophisticated nonlinear dynamic programming problem, we test here the hypothesis that forthe considered task context the brain only needs to learn the unknown taskparameters, composed by arm and object dynamics, in a locally linearsystem identification process. This identification process of task dynamicsenables in turn a decision on the form of the OFC policy. Our approachdescribes motor learning as gradient descent steps in the space ofunknown task dynamics parameters. This mechanism is driven by the errorbetween predicted and actually produced object movements in each trialand can be implemented at the neuronal level by the modification ofsynaptic weights via Hebbian learning rules. The aspect of motoradaptation in our approach involves both the update of task dynamicsand the applied OFC strategy. It thus proposes a novel framework thatexpands and merges studies which on one hand use predictive models ofOFC to fit human performance, assuming that the system dynamics arealready known by humans at the initiation of a task (Nagengast et al,2009) and on the other hand previous work which limits the investigation ofmotor learning to the assumption of updated task dynamics whileconsidering a fixed control strategy throughout the experiment (Berniker &Kording, 2008).
Crucially, our adaptation model predicts accurately the gradualprogression of human learning from trial to trial and performs better incapturing the behavior of subjects at the end of the experiment than anideal observer model, which assumes complete knowledge of task dynamics.Our results suggest that the brain employs simple learning rules to supportdecisions implemented by near optimal control in complex objectmanipulation tasks. Our proposed framework provides thereby analgorithmic formalization, which can guide further experimental investigationson the neural foundation of cortical action selection and motor learningrules.
Keywords: learning, optimal feedback control, action selection
Day 4 (Friday) 10h45
DMB 2014 | 912 Sept 2014 | 87
Franc i s Cr ick Room
TIM RAKOW University of EssexNATHANIEL J. S. ASHBY University of Essex; Carnegie Mellon University
Many choices have consequences for the length of time that it takes tocomplete a task, activity or journey; yet laboratory research on risky choiceis dominated by investigations of the choice between monetary lotteries.Previous research using hypothetical choice scenarios suggests thatpeople’s decisions are less sensitive to temporal expenditures andoutcomes than to monetary ones; suggesting that money is not a goodproxy for time if one wishes to understand decisions about time. In twostudies, we conducted direct tests of decision makers ’ sensitivity tomonetary and temporal outcomes in risky choice. For instance, theparticipant might face a choice between a risky alternative that imposespossible delays of 60 or 2 seconds, and a safer alternative with 31 or 25second delays possible; and his/her preferences would be compared tothose for an equivalent choice involving monetary losses. Choices were nothypothetical; the payoff distributions of each alternative were eitherpresented via descriptions or learned through experience; and outcomeswere framed as either losses or gains. Experiential choices over losses andgains were highly similar for time and money, with preferences developing insimilar ways, and at similar rates, as “good” and “bad” outcomes wereexperienced. The only notable difference between time and money forthese experiencebased choices was a small reversed reflection effect inchoices involving time, with slightly more risk seeking for gains than for losses.Described choices also differed little between the two domains, thoughchoices involving time showed slightly lower sensitivity to changes in framingand expected value. Together, these studies suggest that when theoutcomes of risky choices are consequential then time really is like money.These data challenge several theories which propose that time matters lessthan money when decisions are made; and speaks to the generalizability oflaboratory research using monetary gambles.
Keywords: Risky choice; time delays; time saving decisions; decisions fromexperience; incentivization; lossgain framing
Day 4 (Friday) 10h45
88 | DMB 2014 | 912 Sept 2014
James Watson Room
PETER S. RIEFER University College LondonDANIEL C. RICHARDSON University College LondonBRADLEY C. LOVE University College London
Research in game theory has investigated coordination in severalscenarios, such as the battle of the sexes or marketentry games, butneglected the fact that we often face situations where we coordinate aspart of a large group of people rather than dyads or small parties. Forexample, if we choose which of two routes to take to work, we will considerour experiences and try to anticipate which of the two options will be lessbusy and therefore faster to drive on. Similar to the routes scenario, ourstudy examines the coordination behavior of large groups with about 50people in a zombie game where players have to choose in which of twohiding places to hide from attacking zombies. Points were distributedaccording to how many people were in a certain hiding place, givinghigher rewards to players in less crowded places. In order to investigatehow people use experiences and build anticipations about other people’sbehavior, subjects played the game for several rounds.
We conducted two experiments using this zombie game. In experiment 2,we added volatility information for subjects in one condition who couldthen see how many people switched from one hiding place to the othercompared to the previous round. Using this approach, we tried to examinehow information about other people’s switching influences individual beliefsand decisions in coordination.
Both experiments show that groups generally struggle to coordinateand reach equilibrium, which is at a fiftyfifty split between hiding placesand describes the situation, where no one can do better by switching tothe other hiding place. When groups come close to equilibrium or evenreach it, they will almost certainly drift away from it, due to some individualswho keep on switching choices. Regarding these findings, it seems unlikelyfor large groups to find and retain stable equilibria, since there will mostlikely be some people who nudge the group right back into volatility. Oncloser look, we found that switching choices after unsuccessful rounds ismainly responsible for lower individual scores, whereas switching aftersuccessful choices showed no significant impact on performance. Wefurthermore found that while switching after unsuccessful choices relates tomore myopic thinking, such as choosing the currently better option in theupcoming round, switching after successful choices is anticipatory andrelated to people’s strategy of integrating expectations about others.
Results of experiment 2 replicated the findings of experiment 1. Althoughperformance wasn’t significantly different in the conditions, we observedsome behavioral differences. Subjects with volatility information switchedmore frequently after unsuccessful rounds than successful ones, while thosewithout volatility information didn’t. When examining people’s strategiesindividually with a probabilistic choice model, we found that volatilityinformation seems to make people rely on more myopic strategies. Overall,these findings suggest that seeing the switching behavior of others causespeople to fall back on simple strategies where they immediately changefrom an unsuccessful option to the currently better one.
Keywords: Group behavior, Decisionmaking, Coordination
Day 4 (Friday) 10h45
DMB 2014 | 912 Sept 2014 | 89
Rosa l i nd Frank l i n Room
RAFAL BOGACZ University of Bristol; University of OxfordPAUL HOWARDJONES University of BristolKEVIN LLOYD University of Bristol
Sometimes the best course of action is to do nothing (nonaction).However, it is unclear how the brain learns the value of nonaction and howit is compared to the value of action during decision making. To addressthese questions, we repeatedly confronted human participants with thechoice of selecting one alternative with a button press or the other bydoing nothing. Analysis of the behavioural data with reinforcement learningmodels did not reveal differences between the rate at which participantslearned the values of action and nonaction. At feedback, in addition tobrain activity correlated with the updated value of action and actionprediction error, we found brain activities correlated with the updatedvalue of nonaction and nonaction prediction error. The representations ofnonaction prediction error and nonaction value were found in right andleft inferior frontal gyri, respectively, which are regions previously implicatedin studies of response inhibition. Additionally, we found value comparisonactivity at feedback in dorsomedial frontal cortex and frontopolar cortex,the latter predicting individual differences in exploitative behaviour andperformance across participants. Our results suggest that the processes oflearning the value of nonaction resemble those for learning the value ofaction, engaging response inhibition regions in a manner analogous to theengagement of regions involved in the planning and implementing of actionwhen learning the value of action.
Keywords: decision making, value, response inhibition, fMRI
Day 4 (Friday) 11h30
90 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
ANDREW HOWES University of BirminghamRICHARD L. LEWIS University of MichiganPAUL WARREN University of ManchesterGEORGE FARMER University of Manchester
Computational rationality is an approach to explaining behavior inwhich theories are specified as optimisation problems defined by abounded machine, a utility function and the experience of an adaptationenvironment (Lewis, Howes, & Singh, 2014; Howes, Lewis, & Vera, 2009).Computational rationality holds out the promise of enhancing thecontribution of rational analysis to testing theories of psychologicalmechanism and thereby offers a means of bringing together mechanisticand rationalistic approaches to psychology (Howes, Lewis, & Singh, 2014).It also promises explanations of behaviors that have been widelyconsidered to be irrational but may be seen as rational given boundedneural architectures (Dayan, 2014; Holmes & Cohen, 2014) and boundedexperience (Hahn & Warren, 2009; Le Mens & Denrell, 2011). In this talk Iwill introduce computational rationality and illustrate it with analyses of twokey decision making phenomena: (1) preference reversals, and (2) riskpreferences. I will show these phenomena are computationally rationalconsequences of uncertainty in expected value calculation.
Preference reversals occur when a preference for one choice overanother is reversed by the addition of further choices. It has been arguedthat the occurrence of preference reversals in humans suggests that theyviolate the axioms of rationality and cannot be explained with utilitymaximization theories. In a recent article we used numerical simulation todemonstrate that for a range of types of contextual preference reversal,including the attraction, compromise and similarity effects, the rational,expected value maximizing, choice between existing options is one that isinfluenced by the addition of new choices (Howes, Warren, Farmer, ElDerredy, & Lewis, 2014). The analysis assumes that people rationallyintegrate two sources of information, one based on an estimate ofexpected value and one based on observation of the rank order ofattribute values. I will also show that the same assumptions explain someobserved risk preference effects, particularly risk aversion in the domain ofgains and risk seeking in the domain of losses.
Keywords: Bounded optimality, computational rationality, decision making,preference reversals, risk preference.
Day 4 (Friday) 11h30
DMB 2014 | 912 Sept 2014 | 91
James Watson Room
DAN BANG University of Oxford; Aarhus UniversityLAURENCE AITCHISON University College LondonPETER E. LATHAM University College LondonBAHADOR BAHRAMI Aarhus University; University College LondonCHRISTOPHER SUMMERFIELD University of Oxford
Humans and other animals use estimates about the reliability ofsampled evidence to arbitrate between competing choice alternatives[Kepecs et al., 2008, Nature]. However, unlike other animals, humans canreport on these estimates, saying “ I ’m sure it ’s this one” when the evidence infavour of a specific alternative is deemed highly reliable. Groups ofindividuals use such confidence reports to resolve disagreement aboutwhich choice alternative to select, usually opting for the one supported byhigher confidence [Bahrami et al., 2010, Science]. To do so optimally,individuals must solve a ‘mapping problem’: they must map their ‘ internal ’estimates onto ‘shareable’ confidence reports so as to maximise theprobability that the group selects the better choice alternative [Bang et al.,2014, Concious Cogn]. For instance, if A on average samples more reliableevidence than B, then A should on average be more confident than B. Inthis talk, I report on two experiments in which we addressed the mappingproblem in the context of a visual perceptual task.
In the first experiment, we tested how pairs of individuals solve themapping problem. Two participants (15 pairs in total) privately made abinary decision about a brief stimulus and indicated how confident they feltabout this decision on a scale from 1 to 6. The responses were then shared,and the private decision made with higher confidence was selected as thegroup decision. Feedback about the accuracy of each decision wasdisplayed before the initiation of the next trial. A comparison of modelsindicated that participants ’ strategy for maximising group accuracy was tomatch their distributions over the confidence scale. For participants withsimilar levels of individual accuracy, confidence matching yielded a relativeincrease in group accuracy (cf. they should on average be equallyconfident). In contrast, for participants with dissimilar levels of individualaccuracy, confidence matching led to a relative decrease in groupaccuracy (cf. the more accurate participant should on average be moreconfident than the less accurate participant).
In the second experiment, we tested the adaptive/maladaptive effectsof confidence matching more directly. One naive participant (38participants in total) performed the visual perceptual task together with foursimulated partners in a two (accuracy) x two (confidence) withinsubjectdesign. The simulated partners were less or more accurate and less or moreconfident than the participant. In line with model predictions, confidencematching yielded a relative increase in group accuracy when participantsinteracted with "poorly calibrated" partners ("less accurate but moreconfident" and "more accurate but less confident"), but led to a relativedecrease in group accuracy when participants interacted with “wellcalibrated” partners ("less accurate and less confident" and "more accurateand more confident").
We suggest that confidence matching arose as a default strategy forcooperation in a world where individuals are rarely perfectly calibrated[Fleming et al., 2010, Science]. Crucially, confidence matching iscomputationally inexpensive, works for different tasks and is not dependenton the presence of feedback about individual accuracy.
Keywords: social interaction; confidence; decisionmaking; computational model;perception
Day 4 (Friday) 11h30
92 | DMB 2014 | 912 Sept 2014
Rosa l i nd Frank l i n Room
M.C. KEUKEN Max Planck Institute for Human Cognitive and Brain Sciences;Cognitive Science Center AmsterdamL. VAN MAANEN Cognitive Science Center AmsterdamR. BOGACZ University of OxfordA. SCHÄFER Max Planck Institute for Human Cognitive and Brain SciencesJ. NEUMANN Max Planck Institute for Human Cognitive and Brain SciencesR. TURNER Max Planck Institute for Human Cognitive and Brain SciencesB.U. FORSTMANN Max Planck Institute for Human Cognitive and Brain Sciences;Cognitive Science Center Amsterdam
In our everyday life we constantly have to make decisions betweenmany different choice options. Recently, quantitative mathematical andneurocomputational models have been developed that make predictionsabout brain structures involved in decisionmaking with multiple alternatives.One such model is the Multiple Sequential Probability Ratio Test (MSPRT),which predicts that activity in the subthalamic nucleus (STh), a smallstructure in the basal ganglia, becomes more active with an increasingnumber of choice alternatives.
The present study set out to test this hypothesis using ultrahigh 7Tstructural and functional magnetic resonance imaging in healthy humansubjects. By simulating the MSPRT model, we generated concretepredictions that were then compared to the observed behavior as well asthe activation pattern in the STh. Preliminary results indicate the involvementof the STh in decisionmaking.
Keywords: Multiplechoice decisionmaking, Subthalamic nucleus
Day 4 (Friday) 12h00
DMB 2014 | 912 Sept 2014 | 93
Franc i s Cr ick Room
LAURENCE HUNT University College LondonNISHANTHA MALALASKERA University College LondonDANIEL GRIGAT University College LondonRAY DOLAN University College LondonSTEVE KENNERLEY University College London
A central feature of many realworld decisions is that each alternativeconsists of several attributes. Such multiattribute decisions may be realisedin many different ways ranging from attributebased strategies (such aseliminationbyaspects) to alternativebased strategies (such as weightedadding of attributes). These strategies make opposing predictions as tohow information will be acquired during decision formation, and, importantly,to how neural circuits can implement the decision as it is being made.Information acquisition may also frequently involve costs to the decisionmaker.
In this study, we consider what normative principles might governinformation search strategies in a multiattribute choice task, and whetherthese match with empirical observations in humans and macaque monkeys.Two choice alternatives, consisting of two attributes, were presented.Subjects sequentially selected which feature of each alternative theywished to reveal; they were also able to terminate information samplingearly in order to make a choice. Potential information that might beuncovered at each turn was drawn from a flat probability distribution, withwhich the subjects were familiarised with before commencing the task. Thismeant that a normative dynamic programming approach could beadopted, to derive the optimal strategy for calculating the value ofgathering information at each turn. The dynamic programming modelprovides normative predictions of both when information sampling shouldbe terminated, and also which information is most valuable to sample next, ifreward is to be maximised.
We probed information gathering behaviour of both human andmacaque subjects on analogous versions of the decision task. Human datawas collected from a large subject pool (>8,000 participants) via asmartphone app, and compared to laboratory data from a smaller subjectpool (21 participants, collected whilst undergoing magnetoencephalography). Choice data from two macaque monkeys was collectedwhilst undergoing neurophysiological recording from prefrontal cortex.Whereas human subjects paid explicit costs for sampling information,macaque subjects underwent the opportunity cost of time.
Some key aspects of human and macaque behaviour matched well withpredictions from the normative model. For example, subjects would terminateinformation sampling early if informative cues had been received that madeone alternative much more likely to be rewarding. However, there were alsointriguing and unambiguous violations of the normative model. For example,human subjects were found to be particularly biased toward reducinguncertainty about the value of the currently preferred alternative, even ifother information searches would prove more valuable. Macaque subjectsshowed a similar bias, in that they would often be unwilling to terminate adecision when confronted with one very poor alternative, without firstascertaining information about the other alternative. I will present potentialconsiderations for the origins of these nonnormative behaviours, and whatthey may imply for the implementation of such decisions in neural circuits.
Keywords: Information search, Multiattribute decision making
Day 4 (Friday) 12h00
94 | DMB 2014 | 912 Sept 2014
James Watson Room
REBECCA FLOYD University of BristolDAVID LESLIE University of BristolSIMON FARRELL University of Western AustraliaROLAND BADDELEY University of Bristol
The adage 'two heads are better than one' suggests that when twopeople discuss an ambiguous situation, the outcome of a joint decision willbe closer to reality than each individual's evaluation. It has beensuggested that this has contributed to the success of our species byenabling us to carve out a unique ‘sociocognitive niche’ (Whiten andErdal, 2012). However, research has led to varying opinions about whetherthere is a benefit to working together e.g. 'Wisdom of the Crowd' (Galton,1907) vs Groupthink (Janis, 1971).
Our studies seek to understand the processes underlying group decisionmaking and why they might cause such divergent findings. We useambiguous stimuli presented to pairs of participants. In each trial,participants individually indicate their perception of the correct answer,before then discussing and agreeing a joint answer. Together with pretesting and selfreported personality questionnaires, the joint decisionprocess is examined in the context of relative individual ability at the taskand interpersonal factors.
Our findings so far suggest that at a broad level, the adage holds. Onaverage, the joint answers were closer to correct than the answers ofindividual participants. However, looking at pairwise performance, the jointanswers were often less accurate than those given by the more "sensitive"(i.e. more accurate) participant, especially when there was a largediscrepancy in participant's abilities to detect dot direction. It appearsthat a weighting of individual estimates in the consensus decision reflectsthe precision of the initial estimates and we aim to interpret these results inthe context of a Bayesian model of how the participants combine theirpersonal information.
Keywords: Joint Decision Making, Collaboration
Day 4 (Friday) 12h00
DMB 2014 | 912 Sept 2014 | 95
Rosa l i nd Frank l i n Room
NATHAN LEPORA University of Bristol and Bristol Robotics Laboratory
A main principle underlying animal perception is the accumulation ofevidence for multiple perceptual alternatives until reaching a preset beliefthreshold that triggers a decision, formally related to sequential analysismethods for optimal decision making. In a series of papers, we haveformalized a Bayesian perception approach for robotics based on thisunderstanding of animal perception. Our formalism extends naturally toactive perception, by moving the sensor with a control policy based onevidence received during decision making. Benefits of active Bayesianperception include: (i) robust perception in unstructured environments; (ii)an orderofmagnitude improvement in acuity over passive methods; (iii) ageneral framework for Simultaneous Object Localization and IDentification(SOLID), or perceiving 'where' and 'what'; and (iv) a formalism that naturallyintegrates with reinforcement learning so that both the active control policyand the appropriate belief threshold can be tuned appropriately to thecontextual situation. A strength of the formalism is that it connects closelywith leading work in neuroscience, allowing insights from animal perceptionto be transferred to robot perception. For example, these methods haveenabled the first demonstration of hyperacuity in robot touch, givingperceptual acuity finer than the sensor resolution, as is common in animalperception. As discussed above, they also give robust perception inunstructured environments
in which there is uncertainty in both where and what objects are, as isalso a central aspect of animal perception. In this talk, we describe this bioinspired approach to robot perception and how it connects withcomputational neuroscience, in particular the macroarchitecture of thecortex and basal ganglia.
Keywords: Decision making, robotics, active perception
Day 4 (Friday) 12h30
96 | DMB 2014 | 912 Sept 2014
Franc i s Cr ick Room
GEORGE FARMER University of ManchesterPAUL WARREN University of ManchesterWAEL ELDEREDY University of ManchesterANDREW HOWES University of Birmingham
The attraction effect (Huber, Payne & Puto, 1982) occurs when achoice between two alternatives is biased by the addition of a thirdirrelevant alternative. In a choice between a 70% chance of £20 (thecompetitor) and a 35% chance of £40 (the target), the addition of analternative in which there is a 33% chance of £36 (the decoy) is likely tobias some people toward the target £40 prospect. This demonstration ofcontext sensitivity in human decisionmaking is typically characterised asirrational. This is because rational valuemaximising models of decisionmaking show that the best alternative can only be guaranteed to bechosen if the available alternatives are evaluated independently of oneanother (Luce,1959).
With a mathematical analysis of computationally rational choice wehave shown that when there is uncertainty about the expected value ofeach alternative, the third decoy option is not, in fact, irrelevant. This isbecause the dominance relations provide information about the likelihoodthat either of the original (target or competitor) options will be best. Aconsideration of the rank order of choice probabilities and values leads tothe attraction effect and higher expected value than would be achieved ifthe "irrelevant" alternative were ignored. The ordinal relations in a twoalternative choice set where one alternative has higher probability, and theother has higher value imply approximately equal expected values.However, the ordinal relations necessary to elicit the attraction effect implythe target has greater expected value than the competitor.
The increase in expected value that results from selecting the targetmay be attenuated in two ways. First is simply to increase the difference inexpected value between the target and competitor prospects. Second isto increase the accuracy which the decision maker perceives the expectedvalue of each alternative. We report the results of an experiment in whichwe test this model using two different paradigms, participants either statedwhich prospect they preferred or which rectangle had the larger area. Forprospects, we systematically manipulated the difference in expected value,while for rectangles, we manipulated the difference in area. As thedifferences increased so the attraction effect decreased. The increasedease of the area task relative to the prospect task resulted in a reducedattraction effect and a greater sensitivity to differences in area. Thetendency for people to exhibit the attraction effect is strongest whenexpected value rank ordering is most uncertain, that is precisely when theeffect is most likely to be rational.
Keywords: Attraction Effect; Preference Reversal; Expected Value
Day 4 (Friday) 12h30
DMB 2014 | 912 Sept 2014 | 97
James Watson Room
KARSTEN OLSEN Aarhus University; Aarhus University Hospital
Humans rarely make decisions in social isolation. Rather, we often makedecisions in the context of other individuals, either via observation or ininteraction with them. Previous research has indicated that throughconfidence sharing we can optimally integrate information from others intoour decisions (Bahrami, Olsen et al, 2010), and, over time, build up asubjective estimate of other's performance even without access toobjective feedback (Bahrami, Olsen et al., 2011; Olsen et al, inpreparation). In the present study, we investigated how individuals learnabout others' performance and their decision confidence in a twopersonperceptual decisionmaking task.
We staircased one subject (the observer) and asked them to identify aGabor patch target (and rate their decision confidence) in an orientationdiscrimination task. The other subject (the mentaliser) was shown theupcoming target and asked to guess on each trial the upcoming decisionof the first subject and then rate their confidence in this guess. Nointeraction was permitted at this stage. Using a signal detection theoreticapproach, we subtly manipulated difficulty and payoff contingency(discriminability and bias) in the first subject's (the observer) task.
Results suggest that individuals can learn and track simple perceptualbehaviour across changes in stimulus conditions. That is, individuals couldreliably pick up on the underlying disciminability parameter behind theirpartner's perceptual decision and their decision confidence. In addition, wefound that the 'mentaliser' systematically underestimated the 'observers'behaviour. These results, and interindividual differences, are discussed inlight of the role of social interaction and information sharing.
Keywords: observational learning, social interaction, collective decisionmaking,decision confidence
Day 4 (Friday) 12h30
98 | DMB 2014 | 912 Sept 2014
MARLÈNE ABADIE University of ToulouseLAURENT WAROQUIER University of SussexPATRICE TERRIER University of Toulouse
The unconsciousthought effect occurs when distractionimproves complex decision making. In three experiments, weinvestigated the effect of a detailed or a globalpresentation format of decision information (i) on the qualityof decisions made after distraction, conscious deliberationor immediately and (ii) on memory for decisionrelevantinformation. We used the processdissociation procedure tomeasure recollection and familiarity (Experiment 1 and 2)and the simplified conjointrecognition paradigm todissociate verbatim and gist representations (Experiment 3).Conscious deliberation resulted in better decisions when theformat was detailed whereas distraction improved decisionquality when the format was global. A detailed formatallowed participants to retrieve precise memories as shownby an increase in recollection and verbatim memory. Gistmemory increased after distraction when a global format wasused. This suggests that conscious deliberation efficiency isdependent upon the availability of precise memorieswhereas the unconsciousthought effect is accompanied byenhanced gist memory.
Keywords: Conscious deliberation; Unconscious thought;Distraction; Presentation format; Decision making; Dualmemory processes
SOPHIE MEAKIN University of Bristol
Decisionmaking is an intrinsic part of everyday life, andso being able to mathematically model such behaviour is ofinterest in a range of areas, including psychology, economicsand machine learning. One approach to modelling discretechoice behaviour in psychology is based on Luce’s ChoiceAxiom, from which we can derive the Independence fromIrrelevant Alternatives property. The family of discrete choicemodels used in economics are known as random utilitymodels. Thompson sampling and Boltzmann exploration aretwo heuristics for sequential decision making used in machinelearning that tradeoff between exploration and exploitation.Despite having dramatically different backgrounds, we canfind links between each of the approaches, whichdemonstrate that they are in fact more similar than we firstexpect.
EGLE BUTT Kingston University, LondonDR GAELLE VALLEETOURANGEAU Kingston UniversityDR MARIE JUANCHICH Kingston UniversityPROF FREDERIC VALLEETOURANGEAU Kingston University
This study explored the effect of severity and base rateon the attributed meaning of probability in the light ofpoliteness theory and examined how these factors mayinfluence numerical estimates of a verbal probability phrasein a legal context where a lawyer communicates uncertaintyabout a client ’s imminent conviction. We expected that afrequent conviction rate and a severe court sentence wouldlead to increased preferences for facemanagementinterpretations of the probability term possible uttered by thelawyer, whereas a low conviction rate and a mild courtsentence would lead to a likelihood interpretation. We alsoexpected that a frequent conviction rate and a severe courtsentence would lead to higher probability estimates of theclient ’s guilt. Data showed that base rate directly influencedthe attributed meaning of possible as well as its numericalinterpretation. Participants preferred the hearerfacemanagement interpretation and ascribed higher numericalestimates in high base rate conditions, whereas theypreferred the likelihood communication interpretation andascribed lower numerical estimates in low base rateconditions. Outcome severity led to an increased preferencefor likelihood communication and had no effect on numericalestimates of a verbal probability phrase. These findingssuggest base rate is a strong determinant of both linguisticand numerical interpretations of the probability phrase‘possible’ and might override a severity bias. We discussmethodological implications for future research on outcomeseverity.
Keywords: outcome severity, base rate, verbal probability,politeness theory, facemanagement devices.
TOMAŽ KOLAR University of LjubljanaPETRA DEBENEC University of Ljubljana
Economy and finance heavily relies on assumption ofrationality of economic subjects. Rich body of theoreticaland empirical arguments however attests that at financialdecisions rationality assumption is often and systematicallybreached. According to De Bondt (1998) and other authorsindividual investors are prone to the various decisionmakingand judgment error ’s; they “discover ” naive patterns in pastprice movements, share popular models of value, investmentsare not properly diversified and trade in suboptimal ways.
DMB 2014 | 912 Sept 2014 | 99
Accordingly, one could expect that unexperienced investorsfrom transitional markets like Slovenia are even less "rational"and more prone to various decision errors. Consistent withProspect theory people in general regularly follow mentalshortcuts (heuristics) and are prone to various biases. Inorder to examine some of these biases (i.e. ambiguityaversion, framing, anchoring and mental accountancy) anempirical study was carried out on a convenience sample of140 individual, nonprofessional investors in Slovenia.Obtained results attest that with certain caution due tosample size and structure, we can conclude that Sloveneinvestors are susceptible for three of the examined biases.We found that respondents are indicating aversion toambiguity, since 73.6% of respondents prefer the “knownratio” choice. They are also prone to framing, since positivelyframed choice was chosen by 78.6% respondents. Inaddition they use different mental accountancies, asindicated by different structure of investments decisions for“easy” vs. “ regularlyearned” money. In regard to anchoringbias obtained results are inconclusive, since respondentsevaded estimation of multiplication result (and calculated it).Such response suggests, that respondents prefer certainoutcomes over uncertain and in fact tend to avoidjudgement error (!). This finding is in line with the ambiguityavoidance bias and suggests that when investors havechoice they are trying to operate with known, certainoutcomes. In investment uncertainty is however not an option,but rather “certain sacrifice for uncertain benefit ”. Thequestion is therefore how and to which extent investors canreduce and cope with market uncertainty (and when thistendency becomes biased), especially when it comes toprivate and inexperienced investors. The question of strengthof bias effect is also relevant at the mental accountancybias, which for instance was not found to be present in strong(extreme) form. Respondents namely did not decide to“ lavishly spend” easy earned money entirely, but alsoinvested it into real estate and other long term and stableinvestment options. The question of how investors can furtherreduce their mental accountancy bias and properly diversifytheir investments, however remains a challenging task. In thisrespect coping with the framing bias seems a bit simplermission, as investors might be warned and trained in order toresist “ framed” promotional descriptions of investment options(e.g. when only positive / negative attributes areemphasized). In any case biases reduce quality of investmentdecisions and proper education of investor is warranted,since various studies (e.g. Ginyard, 2001; Dorn & Huberman,2005) attest that knowledge and experiences affect qualityof investment decisions.
Keywords: Behavioral economics, Heuristics, Psychologicalbiases, Investors, Slovenia
JACK DOWIE University of Southern DenmarkMETTE KJER KALTOFT Norwegian Knowledge Centre forthe Health ServicesØYSTEIN EIRING University of Southern DenmarkJESPER BO NIELSEN University of SydneyGLENN SALKELD London School of Hygiene & TropicalMedicine
Personcentred care is the increasingly avowed aim ofhealth services and professionals. To be meaningful suchcare requires a shared decision making process in which anindividual's preferences over the multiple criteria that matterto them are synthesised with the Best Estimates AvailableNow (BEANs), at the point of decision, for how well each ofthe available options will perform on each criterion.Conventional evidencebased approaches can meet thelatter requirement in relation to the required BEANs only byassuming professionals are able to make up the shortfallsremaining after the peerreviewed published products ofscientific research have been fully exploited.
Since the clinical judgement of individual professionalshas never been subjected to scientific validation in thisrespect, we have a situation wheredemonstrated scientific rigour is simultaneously regarded asessential and irrelevant to clinical decision making. Attemptsto increase the external validity of scientific studies (notablyrandomised controlled trials) are attractive to many, but cannever get near to providing most of the BEANs needed inpersoncentred practice.
What is required is the translation of the wisdom of theclinical crowd through the systematic processing of thebeliefs of expert professionals. Given the variations inclinician’s beliefs and limited statistical competencies (asestablished by Gigerenzer and colleagues), these beliefsrequire analytically rigorous processing. Needed therefore isthe reformulation of the research community's task as beingthe continuous (‘ living ’) production of BEANs, within aBayesian framework and by a process that is ARAPAN AsRigorous As Practical And Necessary. This will involve thesystematic elicitation and analysis of expert beliefs, as well asthe exploitation of observational studies and big dataincluded in health records and elsewhere. We explore theuse of Expertisebased Network MetaAnalyses in thiscontext.
Analytic rigour should not be confounded with scientificrigour. A binary concept of ‘evidence’, as that which meets,or doesn't meet, some general, value judgementbasedthreshold (e.g. p<.05), is inappropriate for individual decisionmakers, who have their own tradeoffs and error loss functions.It creates much of the 'knowdo' gap and perceived‘ translation’ problem at the final ‘bedside’ stage.
To give substance to this argument we inserted the resultsfrom a recent highquality Mixed Treatment Comparison(Network MetaAnalysis) on medications for GeneralisedAnxiety Disorder into a MultiCriteria Decision Analysis(http://healthedecisions.org.au/respond/mental). The criteriaincluded (Response, Remission, and Tolerability) were
100 | DMB 2014 | 912 Sept 2014
determined by data availability in trials, not by a survey ofpersonaspatient important outcomes. Minimally the latterwould distinguish different types of side effects and adverseevents, given the known existence of, and differential concernwith, effects on sexual, and other functions. Working to thestandards appropriate to practice as opposed to science,and simultaneously increasing the range of source inputs toinclude expert beliefs, is essential to give such modellingresearch practical relevance for personcentred care. Majorbenefits in establishing priorities for personcentred researchwill follow by way of ‘backward translation’ of the need forbetter BEANs for many personaspatient importantoutcomes.
Keywords: personcentred care; decision making; expertbeliefs; Network MetaAnalysis
BARNEY DOBSON University of BristolJOLYON MILESWILSON University of BristolDAVID LESLIE University of BristolTHORSTEN WAGENER University of BristolJONATHON ROUGIER University of BristolIAIN GILCHRIST University of Bristol
The need for effective public awareness of flood hazardsis becoming increasingly evident, especially in light of theongoing environmental consequences associated withclimate change. Researchers and governmental bodiescurrently emphasise the necessity of publicly available floodrisk information with which members of the public can baseappropriate flood mitigation strategies, but the success ofthese communication efforts appears to vary. Much of thisvariation has been observed to be at least in part due toindividual differences such as previous experience with floodhazards and sociodemographic variables such ashomeownership and level of income. Relatively less attentionhas been directed toward how the visual organisation ofcommunication mediums affects the viewer ’s interpretationsand inferences, though there is some evidence that suchfactors are likely to be influential. Research in the cognitivesciences over the past 20 years has provided severaltheoretical models for how visual displays are perceived,analysed, and decoded by the human visual system, and it ishighly likely that these processes influence interpretation inthe context of risk communication. With this in mind, weinvestigated how the format of presentation floodriskinformation influences viewers ’ decisionmaking and risktakingbehaviour. Participants were presented with pairs of housesand asked to decide which house they would prefer to buy.Information for each house – such as price, energyperformance, etc. – was presented alongside floodriskinformation to encourage participants to consider variousaspects of the house. Within and between participants, wecompared the flood hazard map form of communicationcurrently used by the UK Environment Agency with two newlydevised formats; a table that presents a combination of
flood depth and flood frequency, and a graphic thatdepicts these depthfrequency combinations with referenceto a cartoon house. The results indicated that participantswere more tolerant of high probability floodrisk houses whenthis information was presented in map format, as comparedto the table and graphic formats. Put another way, higherriskhouses were more frequently rejected when the floodriskinformation was presented in the table and graphic formats.These results confirm the suggestion that the way in whichfloodrisk information is presented affects how acceptablethe risk is, not only between participants but also for thesame individual in making a series of decisions. The findingthat flood hazard maps encourage greater risk acceptancethan other types of risk presentation is potentially concerningin light of the fact that this method of communication iscurrently prescribed by the 2007 Flood Directive ofEuropean Parliament. The finding that this variation existseven within the individual is particularly important consideringthe emphasis placed on differences between individuals inthe risk communication literature. It may therefore be useful forcommunicators to direct attention toward both the methodof risk presentation in addition to the intended audiencewhen developing floodrisk communication mediums. Theseearlystage findings highlight a need for greaterconsideration of presentation format in floodriskcommunication, and further research in this area will hopefullycontribute to more effective risk communication in the nearfuture.
Keywords: flood, risk, presentation, communication, visualorganisation, format, maps, environment, public perception,prevention, awareness, psychology,
DR JOHN GOUNTAS Murdoch UniversityDR JOSEPH CIORCIARI Swinburne University of Technology
The dual cognitive processing model has a longtradition and is considered to be robust by a number ofresearchers (Evans and Over, 1996; Sloman, 1996, 2002;Stanovich, 1999; Kahneman, 2002; Evans, 2012). Stanovichand West (2000) summarised the two broad modes ofthinking as system 1 (fast, holistic, heuristic, intuitive, emotional,experiential, implicit, automatic), and system 2 (slow, rational,analytical, reflective, sequential, explicit). However thisconceptualisation has not been universally accepted as areliable theory. A number of researchers have criticised theputative dual cognitive processing systems for its lack ofconceptual coherence (Osman, 2004; Keren and Schul2009). Kruglanski and Gigerenzer (2011) have expressedreservations regarding the validity and reliability of thecurrent dual processing model and proposed that onesystem is possible, using a unified rule of ecologicalrationality for both intuitive and deliberate judgments.Sherman (2006) suggests that there are severe limitationswith the dual system and postulates that a quad model ispotentially more useful. Cools and Van Den Broeck (2007)suggest that there three cognitive styles which are animprovement on the dual system of analyticintuitive model.
DMB 2014 | 912 Sept 2014 | 101
The main weakness of the dual cognitive systems theory is thediversity (lack of consistency) of definitions and disparateattributes to describe each system. Conceptual incoherenceof what constitutes system 1 and 2 is evident in the literaturebecause of the inclusion of incompatible attributes in systems1 & 2, which are qualitatively different. The lack of valid andreliable measures to identify/measure empirically eachsystem’s attributes creates even more uncertainty about datainterpretation and theoretical coherence. Simon’s (1999)idea of invariance in physical phenomena and humanattributes implies that there is probably invariance in humancognitive processing systems and decision making processes.The dual process systems model ignores the researchliterature on learning styles and personality differences. Theresearch literature on personality and learning styles suggeststhat there is more variability in cognitive processing systemsthan the current dual model permits. This paper integratespersonality theories and learning styles and proposes thatthere are four cognitive systems which use qualitativelydifferent types of information processing systems. The fourputative thinking systems are LogicalIdeational, PragmaticPhysical, EmotionAction, and ImaginativeVisualising. All fourthinking systems are hypothesised to be representative ofdistinctive brain processing systems and therefore there arereal physiological networks for each mode of informationprocessing. The paper reports three separate studies. Thefirst study reports the development of the psychometricconstructs and scales for each of the four cognitive systems;the second study reports the findings of the EEG brain scans,which verify the different brain systems; and the third studyreport the fMRI findings which identify more accurately toBOLD activation for each of the four cognitive processingsystems. All three studies support the hypothesis that peoplethink and process information in distinctively different modesand there is neuroscientific evidence to support theexistence of separate brain functions for each thinkingsystem. The empirical survey, EEG and fMRI findings providerobust evidence of scale validity and reliability. Implicationsfor decision preferences and behavioural applications arediscussed.
Keywords: cognitive processing systems, thinking styles,dual processing, scale validation neuroscience testing.
SILVIO ALDROVANDI Birmingham City UniversityGORDON D.A. BROWN University of WarwickALEX M. WOOD University of Stirling
Despite extensive nutrition information campaigns andfood labelling policies (e.g., Burton, Creyer, Kees, & Huggins,2006), the prevalence of harmful effects arising from poordietary choices is on the rise (e.g., Strazzullo, D'Elia, Kandala,& Cappuccio, 2009). The increase in the proportion ofpeople who are either overweight or obese (e.g., Bassett &Perl, 2004) has been mostly associated with foodoverconsumption, over and above unhealthier life style (e.g.,Chandon & Wansink, 2012). It has been often suggestedthat poor food choices and food overconsumption might berooted in cognitive biases that undermine accurate
evaluations of foods and their healthiness (e.g., Wansink, Just,& Payne, 2009). In an effort to determine the cognitivemechanisms behind these biases, we assess the predictionsof rankbased (e.g., Stewart, Chater, & Brown, 2006) andreferencelevel (e.g., Helson, 1964) models of judgment anddecision making in the domain of food evaluation andchoice. We also test whether the principles embodied in themodels can be used to inform and improve consumer ’sdietary choices through social norm interventions that mayreduce the preferences for relatively unhealthy foods.Recently behavioural ‘nudging ’, a method for guiding andinfluencing consumer ’s behaviour by shaping the environmentbut without unduly restricting their freedom of choice (cf.Thaler & Sunstein, 2008), has been implemented within thesocial norms framework (e.g., Agostinelli, Brown, & Miller,1995). Indeed, many behavioural ‘nudges ’ provide socialnorm information and have been found to induce behaviourchange in a variety of contexts such as energy consumption(e.g., Schultz, Nolan, Cialdini, Goldstein, & Griskevicius, 2007)and recycling (e.g., Goldstein, Cialdini, & Griskevicius, 2008).Generally, behavioural nudges based on changingperceptions of social norms often provide information aboutmean levels of behaviour (e.g., average energy consumption)in a social comparison group. However, recent cognitivemodels of judgment suggest that provision of rankbasedinformation (e.g., percentage of other people consuming lessenergy) might tap more directly into natural ways ofprocessing information. Studies 1 and 2 find that people’sevaluations are highly sensitive to the options available inthe decisionmaking context in ways consistent with rankbased models of judgment and decisionmaking. We alsofind that people’s concern about their own consumption ofunhealthy products is predicted not by the quantity ofconsumption, but by how this quantity is believed(inaccurately) to rank within a social comparison distribution(Study 3). In Study 4 we find that rankbased social normfeedback – telling people how they rank within a normativecomparative sample – increases willingness to pay for ahealthy food by over 30%, with greater effects forparticipants who most underestimate their own rankedposition. Meanbased social norm information (i.e. tellingpeople how they differ from the average person), in contrast,had no effect. We conclude that people’s attitudes towardsfood consumption are determined by inaccurate beliefsabout how much other people consume and that rankbasedsocial norms feedback is most effective in addressingpeople’s inaccurate beliefs and can stimulate healthy foodpurchasing.
Keywords: Food perception; Concern about foodconsumption; Social norms; Decision by Sampling; Contexteffects; Healthy eating
102 | DMB 2014 | 912 Sept 2014
TOBIAS ENDRESS University of GloucestershireTONY GEAR University of Gloucestershire
This paper presents the results of a study of the quality ofequity predictions in online communities. The study reviewssecondary data from several special interest communities(Stockjaeger, Sharewise, and Spekunauten) that focus onstock market topics and create group predictions usinggroup decisionmaking approaches. Additionally, primarydata was generated through purposefully selected groups ina controlled environment. The experiments consist of a pilotand a main run. The groups were benchmarked with actualmarket prices as well as with each other. The data wascollected over a period of 10 weeks. The objective of theresearch project as a whole is to assess the quality of thepredictions in terms of accuracy and clarity. One of the mainchallenges in the online decisionmaking process seems tobe that there are often just a few predictions (if any) foreach stock, potentially outdated predictions, and limitedactivity of community members. There might be somemoderation and/or user guidance needed to address theseissues and to gain a critical mass. Another issue is due to thedesign of the communities. The occasional underperformance of online groups might be influenced by aconforming effect and community members' tendency to makeoverly optimistic recommendations and preferences for highbeta shares (i.e., riskier shares). Such a speculation isconsistent with the results from the study in which therecommendations of the existing communities turned out tobe rather optimistic, as well. Through the application of wellknown best practices from the literature, it might be possibleto improve the accuracy of group predictions. Besidesidentifying possible areas for improvement, the studydemonstrated that online groups might be able to deliverpredictions of similar quality to financial experts. The expertsrepresented in the Bloomberg consensus did not provideprice targets with a higher accuracy than the onlineSharewise group. While the study has provided someindications that in certain situations and with careful groupdesign, stock price predictions can be superior to thepredictions of experts, the main experiment indicates a moredifferentiated picture and provided valuable informationabout the underlying decisionmaking process.
Keywords: online community, group decisionmaking,collective intelligence, equity predictions, stock trading
ANYA SKATOVA University of NottinghamCAROLINE LEYGUE University of NottinghamALEXA SPENCE University of NottinghamEAMONN FERGUSON University of Nottingham
Cooperation is central to an effective and fair society.Common evolutionary mechanism like kinship and reciprocitycannot fully explain cooperation with strangers. Theoretically,emotions could play an important role in helping explaincooperation between strangers in everyday life. Anger hasbeen suggested as a primary mechanism that sustainscooperation, specifically when it is possible to act on theangry impulse to punish free rider. Emotions of shame andguilt should lead to repair of damaged social interactions,and should turn the free rider “ to be good to others”.However, the precise consequences of emotions incooperative situations when sanctioning mechanisms are notpresent are unknown.
In our experiment participants (N=118) took part in twoblocks of inverted repeatedinteractions Public GoodsGames, with ten games in each block. After each gameparticipants reported emotions they felt at that moment onLikert scales, including anger, guilt, shame. We analysed howdecisions of other group members affected people’semotional reactions and behaviour. We demonstrateddetrimental effects of anger towards free riding on theoutcomes of interactions in social dilemmas. Theconsequences of anger were not counteracted by theeffects of shame and guilt, experienced by individualsthemselves. We found that strategies that showed the highestlevels of free riding and caused the collapse of cooperationwere not associated with shame or guilt after benefiting atthe expense of others. Further, these profitseeking strategiesretaliated more frequently than others and their retaliationwas fuelled by anger. We found, however, that suchbehaviour did not secure better profits. Instead, strategiesthat withheld the anger impulse and retaliation towards freeriders made better profits through sustaining cooperation inthe group.
Our results in the line with previous findings, demonstratingthat in a repeatedinteraction scenario retaliation and noncooperation does not secure better payoffs. Further, we showthat, at least in no sanctions scenarios, acting upon one’sanger is detrimental to social interactions as it leads tofurther escalation of retaliation. The effects are enhanced bystrategies that seem to play a zerosum individual profitmaking game. They are angry when their profits suffer andsubsequently they retaliate. In addition, they do not showconcern for the public good or for being a free rider (noshame or guilt). On the other hand, those who signal theirprosociality to the group from the start (by using little in thefirst round) and those who show concern for others (byfeeling guilt and shame when they happened to benefit atexpense of others), repair more often and use more lenient,forgiving strategies. However, our data is at odds withprevious assertions that these individuals are not angry atfree riders. On contrary, we demonstrated both strategieswere associated with anger after being treated unfairly,however not everybody retaliated in response. The strategy
DMB 2014 | 912 Sept 2014 | 103
of not acting upon one’s anger, nonretaliation and repairseem to be more profitable in the long run. We showed thatduring repeated interactions around public good it pays offto control one’s anger and turn the other cheek.
JAN K. WOIKE Max Planck Institute for HumanDevelopmentSEBASTIAN HAFENBRÄDL University of Lausanne
Faced with a decision that influences yourself and others,it has been argued that your choice might heavily dependon the fact whether you know your own position as arecipient or not. While Rawls argues that decision makersbehind a “veil of ignorance” will adhere to the maximinprinciple, Harsanyi proposes the maximization of expectedvalue (and total social welfare), instead.
We conducted a series of threeplayer moneysharinggames, in which participants had to choose betweendifferent monetary distributions for three receiving positions.The distributions varied systematically in terms of achievedequality, total social welfare, and expected personal profit.In study 1 (N=154), one group was asked to choose thefairest distribution in hypothetical games, while otherparticipants made choices for other players and incentivizedchoices under different degrees of ignorance concerningtheir receiver position.
When participants were disaffected by their choices orall three positions were equiprobable, choices were mostlyaligned with choices in the hypothetical fairness condition:Distributions maximizing equality of outcomes were preferredover distributions maximizing average payoffs or the outcomefor specific positions. When participants knew their receiverposition for certain, the majority of participants chose tomaximize their personal outcome. Finally, when participantswere informed that they would be assigned one out of two ofthe three positions (with a chance of p=¾ for one of thetwo), choices were split into three clusters: some participantschose equal distributions, some chose to maximize theminimum payment for the two receivers and a third groupmaximized their expected outcome. Further analyses includereactiontime and qualitative data, and behavior in othereconomic games.
In Study 2 (N=111), we tested participants with sets ofchoices, where the equally distributed option was dominatedby all other options. Indeed, some participants chose aguaranteed loss of CHF 1 for everyone over an averageoutcome of CHF 5 with an unequal split.
In summary, some variants of the veil of ignorance areable to achieve distributive outcomes that are consideredfair by neutral observers. While social welfare maximization isnot pursued by a large number of subjects, a substantialpercentage chooses according to the maximin principleproposed by Rawls. Also, a nonnegligible percentage ofparticipants are willing to incur both personal and socialwelfare costs to achieve equality.
Keywords: veil of ignorance, distributive fairness, groupdecisionmaking
JERYL L. MUMPOWER Texas A&M UniversityGARY H. MCCLELLAND University of Colorado Boulder
The distinction between idiographic and nomotheticapproaches in psychology is widely attributed to Allport(1937). Idiographic research in JDM generally involvesintensive study of a single individual to achieve anunderstanding of that particular individual. For example,GonzalezVallejo et. al (1998) developed separate modelsfor 32 individual physicians ’ diagnostic judgments andtreatment decisions for acute otitis media in children. Suchmodels permit predictions about how individuals will behavewith respect to comparable cases in the future. Comparisonsand contrasts can be made among individuals or clusters ofindividuals regarding their judgment processes or policies.
Nomothetic research in JDM generally involvesinvestigations of large groups of people to find general lawsof behavior that hold for people in general. An example isTversky and Kahneman’s (1981) studies of the effects offraming on preference reversal using the famous socalled“Asian disease” problem. The various cognitive heuristics andbiases identified by Tverksy, Kahneman, and others aregenerally presumed to hold to some degree for all persons.
Typically, idiographic research relies on a singleindividualmany cases paradigm, whereas nomotheticresearch relies on a many individualsone case paradigm.These are not the only design possibilities for judgment anddecision making research, however.
A third possibility is the single decision makersinglecase, or case study, paradigm. A wellknown example of thecase study approach in JDM is Allison’s (1971) study ofdecision making within the U.S. government during the CubanMissile Crisis. Case studies are similar to nomotheticapproaches in so far as they focus on particular decisionmakers, but the validity of generalizing to future cases isunclear since these studies are based on an n of one.
A fourth variant – and the focus of this paper – is themany individualsmany cases paradigm which yieldsdescriptions or models of the judgment and decision makingbehaviors of systems. For instance, Mumpower andMcClelland (2014) analyzed decision making processesduring the referral and substantiation stages of the childwelfare system. The data for the study represented theaggregation of literally millions of decisions by individualsacross literally millions of cases. Analyzing those data withJDM techniques yields “asif ” models of system behavior thatdo not apply either to any individual decision maker or to“people in general”. In Mumpower and McClelland (2014),the systemic, many individualsmany cases, approachuncovered quite different patterns of JDM behavior withrespect to the treatment of different racial and ethnic groupsin the referral and substantiation portions of the child welfareservices system.
In this paper, we review and discuss the Mumpower andMcClelland study, as well as others among the relatively smallnumber of JDM studies that employ the systemic, manyindividualsmany cases paradigm. We conclude that the useof JDM methods to analyze system behaviors has greatpotential for improving our understanding of those systemsfrom both fundamental and applied perspectives.
Keywords: system models, idiographic, nomothetic, casestudies
104 | DMB 2014 | 912 Sept 2014
WOUTER BOEKEL University of AmsterdamERICJAN WAGENMAKERS University of AmsterdamLUAM BELAY University of AmsterdamJOSINE VERHAGEN University of AmsterdamSCOTT BROWN University of NewcastleBIRTE FORSTMANN University of Amsterdam
A recent ‘crisis of confidence’ has emerged in theempirical sciences (Pashler & Wagenmakers, 2012). Recentstudies have suggested that questionable researchpractices (QRPs; Simmons, Nelson, & Simonsohn, 2011) suchas posthoc storytelling and nonpublication may becommon in the field of psychology (John, Loewenstein, &Prelec, 2012). These QRPs can result in a high amount offalsepositive findings, decreasing reliability and replicabilityof research findings. A potential solution is to preregisterexperiments prior to acquisition or analysis of data(Wagenmakers, Wetzels, Borsboom, van der Maas, & Kievit,2012; Chambers, 2013). In this study we attempt to replicatestudies that employ the commonly used approach of relatingbrain structure to behavior and cognition. These structuralbrainbehavior correlations (SBBCs) occasionally receivegreat scientific and media attention. Given the impact ofthese studies, it is important to investigate their replicability.Here, we attempt to replicate five SBBC studies comprising atotal of 17 effects. To prevent QRPs such as posthocstorytelling and nonpublication, we employed apreregistered, purely confirmatory replication approach. Wewere unable to successfully replicate any of the 17 effects.For all of the 17 findings under scrutiny, Bayesian hypothesistests indicated evidence in favor of the null hypothesis. Theextent of this support ranged from anecdotal (BF < 3) tostrong (BF > 10). To our knowledge, this is the first multistudyconfirmatory replication of SBBCs. With this study, we hope toencourage other researchers to undertake similar replicationattempts.
Keywords: preregistration; confirmatory; replication; brainbehavior correlations
CRAIG HEDGE Cardiff UniversityPETROC SUMNER Cardiff University
Behavioural tasks which measure an individual ’s ability toinhibit unwanted response are used prominently in theliterature to examine the mechanisms underlying cognitivecontrol, decision making processes, and in the examination ofindividual differences and psychopathology. Theinterchangeable use of different tasks, and in some cases ofdifferent performance measures from the same task, entailsseveral critical assumptions: First, that the different measures
tap in to the same underlying mechanisms; and second, thatthese measures reflect an underlying trait of the individualthat is stable over time. Despite the importance of theseassumptions, little evidence exists to support them for manywidelyused tasks. Measures of response control from differenttasks often show modest or inconsistent correlations, andretest reliability across multiple testing sessions is notcommonly examined.
Our study seeks to address this gap in the literature, andprovide guidance on the optimal use of these tasks. Weexamined four common measures of response control; theStroop task, the Eriksen flanker task, the Go/Nogo task, theStopSignal task. In addition, participants completed a selfreport measure of impulsive behaviour. Fortysevenparticipants completed the tasks in two sessions, takingplace three weeks apart. We report data on several keypoints. First, does each of these tasks provide an assessmentof a trait that is stable over time, and second, how manytrials are required for this to be achieved? Third, if thestability of these individual measures can be verified, is therethen a relationship within (e.g., comparing reaction time anderror rates from the same task) and between (e.g., Stroopreaction time cost and Flanker reaction time cost) thedifferent response control measures?
Retest reliability for the various measures ranged from fairto good, and we use a method of subsampling different trialnumbers to illustrate the number of trials required to achievea given level of reliability. However, even with higher trialnumbers, the correlations between the different responsecontrol measures within and between tasks ranged from lowto moderate, prompting caution in assuming that suchmeasures can be used interchangeably. We discuss theimplications of these findings for the use of these tasks inboth theoretical and applied contexts.
Keywords: Response inhibition, Cognitive control, reliability,impulsivity
TIMOTHY BALLARD University of BristolSTEPHAN LEWANDOWSKY University of Bristol; University ofWestern AustraliaMARK HURLSTONE University of Western Australia
Uncertainty is an inherent feature of any climate changemodel. From a normative perspective, greater uncertaintyincreases the risk associated with the impact of climatechange. However, uncertainty is frequently cited in politicaland public discourse as a reason to delay mitigating action.This failure to understand the implications of uncertainty mayhave considerable costs both for the climate and the globaleconomy. It is therefore critical to communicate uncertainty ina way that better calibrates people’s risk perceptions withthe projected impact of climate change.
We conducted an experiment that examined whetherpeople’s concern about projected climate changeoutcomes was influenced by the manner in which uncertaintyis expressed. Specifically, we examined whether concern
DMB 2014 | 912 Sept 2014 | 105
about projected outcomes depended on whether outcomeswere expressed as certain in magnitude but arriving at anuncertain time or uncertain in magnitude but arriving at acertain time. We presented participants with a series ofstatements and graphs indicating projected increases intemperature, sea levels, ocean acidification, and adecrease in artic sea ice. In the uncertain magnitudecondition, the statements and graphs reported the upperand lower confidence bounds of the projected magnitudeand the mean projected time of arrival. In the uncertain timeof arrival condition, they reported the upper and lowerconfidence bounds of the projected time of arrival and themean projected magnitude.
Preliminary analysis suggests that participants in theuncertain magnitude condition reported greater concernabout the projected outcomes than participants in theuncertain time of arrival condition. This result suggests thatthe manner in which uncertainty concerning the impact ofclimate change is communicated affects people’sinterpretation of the information. This finding has importantimplications for effectively communicating the risks associatedwith climate change to policy makers and the general public.
Keywords: Uncertainty, Risk, Judgement, Climate Change
MATTHEW COLLETT University of Exeter
Animals use information from multiple sources in order tonavigate between goals. Ants such as Cataglyphis fortis usean odometer and a sunbased compass to provide input forpath integration (PI). They also use configurations of visualfeatures to learn both goal locations and habitual routes tothe goals. Information is not combined into a unifiedrepresentation (Collett, et al, 1998; Collett, et al, 2003;Wehner et al, 2006; Collett & Collett, 2009), but appears tobe exploited by separate expert guidance systems (Collett,2010; Cruse & Wehner, 2011). Visual and PI goal memoriesare acquired rapidly and provide the consistency for routememories to be formed. Do established route memories thensuppress the guidance from PI? A series of manipulationsputting PI and route memories into varying levels of conflictfound that ants follow compromise trajectories (Collett,2012) The guidance systems are therefore active togetherand share the control of behaviour. A simple model showsthat observed patterns of control could arise from asuperposition of the output commands from the guidancesystems, potentially approximating Bayesian inference (Ma, etal, 2006). These results help show how an insect ’s relativelysimple decisionmaking can produce navigation that isreliable and efficient and that also adapts to changingdemands.
Keywords: Insect decision making, navigation, integration
DAVE E.W. MALLPRESS University of BristolALASDAIR I. HOUSTON University of BristolBENJAMIN Y. HAYDEN University of Rochester
The Balloon Analogue Risk Task (BART, Lejuez et al.2002) is a recent, but hugely influential and wellvalidatedhuman behavioural task that correlates with measures ofimpulsivity, selfcontrol and risk attitude. Subjects performingthe task inflate a virtual balloon that increases in reward sizewith time, but also brings greater risk of bursting, which wouldresult in no reward. The subject must then decide the bestmoment to cease inflating and accept the reward. Here wewill provide a discussion of the strategies that should beemployed by an optimal decisionmaker as well as theconceptual issues with these personality constructs. Severalnovel versions of this task were developed and performed onRhesus macaques to test its worth as an impulsivity measureacross species. Our continuoustime sequential choiceversions of this task share several features with the naturaldecision problems faced by animal foragers, particularly intrading off delay and risk with reward magnitude and ourresults of our experiments show that the monkeys take theballoon reward earlier than they should if behaving optimally.Finally, we will discuss why mathematically equivalent versionsof the task do not necessarily elicit the same behaviour.
Keywords: Impulsivity, Risk, Balloon Analogue Risk Task,Optimality
REBECCA PIKE University of BristolALASDAIR HOUSTON University of BristolJOHN MCNAMARA University of Bristol
Monogamy is the prevailing mating system in birds andmuch is to be understood of the functions and mechanismsbehind these long term affiliations.
In species where incubation is shared by both parents,the mate must make a decision of how long to wait at thenest for their partner to return from foraging e.g. A breedingAlbatross pair.
We modelled a situation with two types of individual inthe population; some individuals will return to the nestwhereas others will abandon it. We identified the incubatingmate’s optimal time to wait until abandoning the nestthemselves given the initial estimate from the population ofthe likelihood their partner is to return, p.
The optimal wait time was found by maximising theincubating mate’s payoff function. As p increases, the optimalwaiting time is longer.
We then considered how gaining information about apartner affected the estimation of p. Differentiating the
106 | DMB 2014 | 912 Sept 2014
optimal payoff function with respect to p allows us toquantify the value of information.
Gaining knowledge allows refinement of your estimate ofp altering your waiting time, thereby increasing the expectedpayoff. The value of knowing your partner may explain whysome species form long term relationships
Keywords: Pair bond, Trust, Value of Information, StatisticalDecision Theory
ANJALI MAZUMDER University of WarwickJAMES Q. SMITH University of Warwick
The case assessment and interpretation frameworkcommonly used by forensic scientists in the UK involves ahierarchy of propositions to be addressed in casework.Forensic scientists are often focused on addressing source orsubsource level propositions, e.g. what is the source of theDNA sample or glass fragment. However, there is increasedinterest in addressing activity level propositions (e.g. did Mr Ybreak the window) to better assist the courts (without strayingoutside the bounds of scientific knowledge). The pairing,framing and development of at least two competingpropositions must be done with careful consideration of thecase circumstances and evidence.
There is a plethora of literature on the use of BayesianNetworks (BNs) for forensic science which expresses thegraphical and probabilistic relationship between measuredvariables. BNs have been particularly useful in providing agraphical representation of the problem, calculatingmarginal and conditional probabilities of interest, andmaking inferences particularly addressing lower levelpropositions. To address activity level propositions, there is aneed to account for different plausible explanations ofsuspect/perpetrator ’s actions and events as it relates to theevidence. In this talk, we propose the use of another class ofgraphical models, chain event graphs (CEGs), exploitingevent tree structures to depict the unfolding of events aspostulated by each side (defence and prosecution) anddiffering explanations/scenarios. Different scenarios canintroduce different sets of relevant information affecting thedependence relationship between variables and symmetry ofthe structure. CEGs are a flexible class of graphical modelswhich can model the asymmetric story structure directly in itstopology. Yet because of its graph modular structure it alsoinherits many benefits of the BN. A BN can be representedas a symmetric CEG but the BN is not always a rich enoughstructure to incorporate all obtainable information.
We demonstrate how CEGs can be very useful inaddressing activity level propositions and assist the courts bydirectly supporting the barrister ’s argument within thetopology of a graph, particularly in complex cases. Itsstructure provides a graphical and probabilistic frameworkfor decisionmaking regarding the validity of a set ofproposed explanations in a systematic way. With the use ofcase examples involving transfer and persistence and
different evidence types, we further show how CEGs canassist in the careful pairing and development of propositionsand analysis of the evidence by addressing the uncertaintyand asymmetric unfolding of the events to better assist thecourts.
Keywords: Bayesian networks, chain event graphs, evidenceanalysis, asymmetry, uncertainty
DMB 2014 | 912 Sept 2014 | 107
NICOLA BOWN Leeds University Business School
Globally, we consume far less food than we produce.Estimates suggest that annually up to half of food producedis wasted, around 1.2 to 2 billion metric tons (Institute ofMechanical Engineers, 2013). In the UK, food wastedannually in the home that could have been eaten isestimated at 4.2 million tonnes, worth £12.5 billion (Wasteand Resource Action Programme, WRAP, 2013). This is aglobal issue that is financially costly and environmentallydamaging. This paper considers how our understanding ofhuman judgment and decision making (JDM) processes cancontribute to understanding and addressing this issue.
Recent UK governmentbacked campaigns such as LoveFood Hate Waste have reported a significant 21% reductionin the amount of useable food that is wasted over the last 5years (WRAP, 2013). What is unclear is the extent to whichthis is attributable to the campaign effectiveness, consumers ’increasing financial constraints or other factors. Although thistopic receives ongoing attention from practitioners andpolicy makers, relatively little is still known about the drivers ofconsumers ’ foodrelated perceptions or the full range ofeffective strategies for communicating about food waste. Thisis due in part to the highly complex nature of consumers ’decisions around food, which include cognitive, emotional,cultural, healthrelated, financial, and moral dimensions(Evans, 2011; 2012). The current paper integrates evidenceof food waste production and reduction interventions (WRAP,2013; Quested et al, in press) with a review of relevant JDMliterature, to provide insight into human psychologicalprocesses during the food planning, shopping, preparation,consumption and storage cycle.
I discuss JDM concepts relevant to household foodplanning and how these may contribute to increased waste,including focalism (usually discussed in relation to theplanning fallacy, Buehler, Griffin & Peetz, 2010; Kahneman &Tversky, 1979; Lovallo & Kahneman, 2003); intertemporalpreferences (Read, Loewenstein & Kalyanaraman, 1999) andvariety seeking (Read & Loewenstein, 1995; Read,Loewenstein, & Rabin, 1999; Simonson, 1990; Simonson &Winer, 1992). We also consider that those responsible forfood shopping and preparation may be willing to pay forhaving a choice of food in order to meet individuals ’preferences. In other words, some consumers will accept alevel of food waste in the home.
An important belief in JDM literature is that humansdemonstrate two types of thinking: System 1 and System 2(Frederick & Kahneman, 2002; Kahneman, 2011). System 1 isfast, intuitive, heuristic, prone to biases, but often functional,and System 2 is slower, effortful, more cognitively demanding,but generally resulting in accurate judgments. Manyrecommendations from waste reduction campaigns (e.g.,“plan what you will eat ”, “and buy what you need”) requirethe consumer to rely heavily on System 2 processes. Whileevidence suggests these are relatively successful strategiesin reducing waste generation (WRAP, 2013), we discuss asyet unexplored System 1 strategies which could contribute to
campaigns incorporating elements of both System 1 andSystem 2, possible to be most effective in achievingsustained behaviour change.
Keywords: Food waste, judgment and decision Making,System 1 and System 2, behaviour change
HANNAH JULIENNE University of BristolDAVID LESLIE University of BristolJAMES HODGE University of Bristol
Dopamine is thought to play a crucial role inreinforcement learning, namely through signalling thedifference between expected reward and actual reward(Schultz et al., 1998). There is increasing evidence that itmay also play a role during decision making, but its preciserole here is less clear. Parkinson’s disease patients show arange of symptoms resulting from dopamine neurondegeneration, including disrupted performance in somereinforcement learning tasks. Interestingly, patients onmedication to increase dopamine levels also show impairedperformance in certain tasks (Frank et al., 2004), furtherindicating a multifaceted role for this neurotransmitter in thelearning and memory process.
The fruit fly Drosophila melanogaster is an ideal modelorganism for dissociating these different roles for dopamine.The fly dopaminergic system serves similar functions as that ofmammals, including a crucial role in reinforcement learning.Reinforcement learning in flies can be assessed using anolfactory assay where flies are trained to associate anodour with a shock punishment or sugar reward. A number offly models of Parkinson’s disease have been developed thatdisplay selective dopamine neuron degeneration. Finally, arange of genetic tools have been developed that allow usto selectively increase or block neurotransmitter release fromspecific subsets of dopamine neurons.
We are taking both experimental and theoreticalapproaches to answering this question. We are testingreinforcement learning and memory in Drosophila models ofParkinson’s disease and comparing their performance to thatof wildtype controls. We are also using thermogenetic toolsto facilitate or inhibit dopamine release at specific timepoints in the learning and decision making processes. We willfit data generated from this to mathematical models ofreinforcement learning and decision making. We hope todetect differential effects of dopamine at memoryacquisition, consolidation and retrieval.
Keywords: dopamine, reinforcement learning, drosophila,Parkinson's disease
108 | DMB 2014 | 912 Sept 2014
GEOFFREY MÉGARDON Cardiff University; AixMarseilleUniversitéeCHRISTOPHE TANDONNET AixMarseille Universitée;Université de GenèvePETROC SUMNER Cardiff UniversityALAIN GUILLAUME AixMarseille Universitée; New YorkUniversity
Topographically organized structures are ubiquitous insensory and motor regions on the brain. These neural mapshave long been modelled with neural fields that incorporatelateral excitation of neighbouring units and lateral inhibitionof remote units (a pattern of lateral interactions often called'Mexican Hat'). Here our aim was to get further insight into thecomputational properties of this organization by testingwhether and under what circumstances it could result in autoinhibition, i.e. whether activity induced by some stimulus sizeswould inhibit itself. Such an explanation has been suggestedfor eye movement data where increasing the size of adistractor can decrease its effect (a 'reversal pattern';Tandonnet et al. J Vis 2012). We show here that a onelayermodel of spiking neurons (a simplification of layers in superiorcolliculus, for example) containing only Mexican Hatinteractions of moderate inhibition extent is sufficient to showautoinhibition of stimuli. This autoinhibition shapes threedistinct activity patterns over time and space according tothe stimulus ’ size. We also observe that a similar reversalpattern to that observed by Tandonnet et al. can beobtained in this kind of model, but only with an ellipticshaped Mexican hat. Interpretations of this elliptic shape arediscussed and invite to further investigation. These resultssuggest that autoinhibition is a possible phenomenon at thelevel of a single neuronal map and that it could participatein the bottomup processing of sensory signals and theinteractions of motor plans.
STEPHEN HEAP University of JyväskyläTAPIO MAPPES University of JyväskyläMIKAEL PUURTINEN University of Jyväskylä
Human social evolution is intricately tied to cooperationand conflict at the levels of the individual and the group.Specifically, humans have a tendency to cooperate withothers at personal cost when cooperation improves theperformance of their own group relative to other groups.However, the relationship between an individual ’s coalitionalpsychology (which forms the proximate basis for socialdecisions) and the emergence of necessary conditions forcooperation to be ultimately selected for remain unclear. Weaimed to experimentally investigate this relationship using a
public goods game, in which participants were repeatedlydivided into random groups of four and presented with asocial dilemma between costly cooperation with their groupand maximising their individual payoff. Additionally,participants could vote on the manner in which their groupinteracted with a randomly paired opponent group. Thechoice of betweengroup interaction determined whetherfunds were competitively taken or cooperatively sharedbetween groups, or whether groups operated in isolationfrom one another. We show that groups of randomlyassembled individuals spontaneously engage in costly groupcompetition, and that decisions promoting betweengroupconflict are associated with high levels of withingroupcooperation. Furthermore, we found that individuals weremore likely to vote for competitive interactions with othergroups as the variance in cooperative behaviour within theirown group decreased and variance in public goodprovisions between groups increased. This result conforms tomultilevel selection theory, which states that betweengroupvariation in fitness must outweigh withingroup variation inorder for individually costly, but group beneficial, traits to beselected for. Additionally, these results support the theoreticalimplication that cooperative actions within groups coincidewith competitive interactions between groups. We analysedthe decisionmaking processes of individuals using statisticaldecision theory to determine how participants utilisedinformation from a dynamic social environment and their ownexperience in order to modify their behaviour in anticipationof future conditions. We discuss these results in terms of howthe nature of information processing corresponds with theultimate drivers of human social behaviour and organisationin an evolutionary context.
Keywords: cooperation, evolution, publicgoods game,ecology
JANINA A. HOFFMANN University of BaselBETTINA VON HELVERSEN University of BaselJÖRG RIESKAMP University of Basel
Making accurate judgments is an essential skill ineveryday life. However, although the relation of differentmemory abilities to categorization and judgment processeshas been hotly debated, the question is far from resolved.We contribute to the solution by investigating how individualdifferences in memory abilities affect judgment performancein two tasks that induce rulebased or exemplarbasedjudgment strategies. In a study with 279 participants, weinvestigated how working memory, episodic memory, andimplicit memory affect judgment accuracy and strategy use.As predicted, participants switched strategies between tasks.Furthermore, structural equation modeling showed that theability to solve rulebased tasks was predicted by workingmemory, whereas episodic memory predicted judgmentaccuracy in the exemplarbased task. Last, the probability ofchoosing an exemplarbased strategy was related to betterepisodic memory, but strategy selection was unrelated toworking memory capacity. In sum, our results suggest that
DMB 2014 | 912 Sept 2014 | 109
different memory abilities are essential for successfullyadopting different judgment strategies.
Keywords: Judgment; working memory; episodic memory;rulebased and exemplarbased processes
JAMES TRIPP University of WarwickADAM SANBORN University of WarwickNEIL STEWART University of WarwickTAKAO NOGUCHI University of Warwick
People often misjudge the probability of the combinationof two events, famously with the conjunction and disjunctionfallacies. Classical probability theory tells us that theconjunction of two events cannot be more probable thaneither event on its own, and the disjunction of two eventscannot be less probable than either event on its own. A longline of research shows us that people tend to violate both ofthese rules, and do so in a way that suggests that they areaveraging the probabilities of the two events instead ofcombining them correctly.
Here we adopt a novel approach to explaining whypeople make these errors. We investigate the possibility thatpeople consider multiple strategies when judgingconjunctions and disjunctions, and use a statisticallysophisticated procedure to combine these strategies:weighting each strategy according to the probability it iscorrect. For example, participants may consider both thenormative combination rule and the mean of the twoprobabilities. Then the strategies are combined in aweighted average using weights that depend on thequestion asked. Our findings suggest that people doconsider multiple strategies when judging conjunctions anddisjunctions.
In Experiment 1 participants estimated how many peopleout of 100 had either (a) one of two attributes, (b) aconjunction of two attributes, or (c) a disjunction of twoattributes. The attributes were denoted by randomly chosenletters and, following Wyer (1976), the probability of anattribute occurring was either "usually", "sometimes", or "rarely",corresponding to a high, moderate, and low probabilityrespectively. This approach allows us to elicit the subjectiveprobability of each adverb for a participant and calculatethe corresponding predictions of each strategy for aconjunction and disjunction. These subjective probabilities ofeach adverb and the weights for each strategy were inferredfrom the data at an individual level.
Our results appear inconsistent with a single strategyaccount. When the adverbs are different, participantresponses to conjunctions and disjunctive lie between theirresponses to each event occurring on its own, producing onaverage conjunction and disjunction fallacies and isconsistent with an averaging account. However, when theadverbs are the same, responses to the conjunction anddisjunction questions were more normative. Taken together,our results show that a single strategy explanation strugglesto explain participant responses and questiondependentweights better explain the data. We further explored why
participants choose the weights they do using moretransparent stimuli and rewards. We presented participantswith events, using spinners, where the correlation betweenevents is clear. Rewards were used to encourageparticipants to use the strategy they considered most likely.The results are discussed in terms of support for single ormultiple strategy of information integration.
Keywords: Conjunction fallacy, disjunction fallacy, informationintegration, probability
GILLES DE HOLLANDER University of AmsterdamERICJAN WAGENMAKERS University of AmsterdamLOURENS WALDORP University of AmsterdamBIRTE FORSTMANN University of Amsterdam
Traditionally, fMRI data are analyzed using statisticalparametric mapping approaches. Regardless of the precisethresholding procedure, these approaches ultimately dividethe brain in regions that do or do not differ significantlyacross experimental conditions. This binary classificationscheme fosters the socalled imager's fallacy (Henson, 2005),where researchers prematurely conclude that region A isselectively involved in a certain cognitive task becauseactivity in that region reaches statistical significance andactivity in region B does not. For such a conclusion to bestatistically valid, however, a test on the differences inactivation across these two regions is required (Nieuwenhuiset al, 2011). Here we propose a simple GLMbased methodthat defines a third ``inbetween" category of brain regionsthat are neither significantly active nor inactive, but rather ``inlimbo". For regions that are in limbo, the activation pattern isinconclusive: it does not differ significantly from baseline, butneither does it differ significantly from regions that do showsignificant changes from baseline. This pattern indicates thatmeasurement was insufficiently precise. By directly testingdifferences in activation, our procedure helps reduce theimpact of the imager's fallacy. The method is illustrated usingconcrete examples.
Keywords: functional magnetic resonance imaging (fMRI),statistical parametric mapping (SPM), imager's fallacy,interaction, univariate modeling, sandwich estimator
110 | DMB 2014 | 912 Sept 2014
TOM STAFFORD University of SheffieldÇIGIR KALFAOGLU Eastern Mediterranean UniversityCOLIN BANNARD University of Texas at AustinLIZ MILNE University of Sheffield
Errors in decision making can reveal the mechanisms bywhich correct performance is normally generated. We presentthree studies of errors in typing and discuss them in terms ofhow goaldirected and habitual processes coordinate indecision making.
Many studies of decision making measure responses in aseries of discrete trials. Such paradigms bring an admiralclarity of what and when the decision under study is. Typing,as a decision making paradigm, sacrifices this clarity multiple decisions are made per second, at multiple levels ofdescription (e.g. word and letter choice) with overlappingtimescales of execution. But even within discrete trial tasks,crossover effects between notionally independent decisionsare found (e.g. negative priming). The advantage of typing isthat it is a realworld task for which highly skilled participantscan rapidly generate a stream of data (keypress identitiesand timings). This data has the advantage of clearly definederrors. For our present purposes we also wish to highlight thattyping is an example of a behaviour which combines bothhighly practised habitual or automatic elements withdeliberate, controlled or goaldirected, action.
Our first study (Kalfaoglu, & Stafford, 2014) involvedasking 19 touchtypists to copytype 100 sentences withoutvisual feedback from either their hands or the screen.Contrary to previous results using discrete trial tasks, wefound no speeding or slowing of decisions prior to errors.Instead we found increased variability in interkey intervals(IKIs) predicted error commission. This heightened variabilitywas found for errors which were corrected (i.e. using thedelete key) and errors which went uncorrected, suggestingthat it was independent of goaldirected errormonitoringprocesses (unlike, for example, error slowing, whichdifferentiated between corrected and uncorrected errors).
Our second study analysed EEG data recorded duringthe first study. Time frequency and Event Related Potentialanalysis were applied the components identified by anIndependent Components Analysis. This analysis showed thatelectrophysiological signals normally associated with errordetection are reliably associated with uncorrected errors.We conclude that they are insufficient to trigger errorcorrection on their own. Synchronisation of activity in mediofrontal regions in the theta range (47hz) appearednecessary for errors to be corrected. This synchronisationalso predicted greater error slowing, demonstrating the coextensive nature of performance monitoring processes anddecision making, in the domain of typing at least.
Finally, we conducted a computational linguistic analysisof error types produced during a second copytypingexperiment (Bannard, Kalfaoglu, Stafford, unpublishedanalysis). Errors were coded as simple motoric errors (e.g.pressing the key adjacent to the correct key) vs "action slips"(which involved the intrusion of some habitual element intoongoing goaldirected action; e.g. intending to type THINKbut actually typing THING, derailing from the correct TH into
the highfrequency ING stem). Our analyses shows howtransitional probabilities derived from a large corpus ofwritten English show significant differences between the threedecision outcome categories. We discuss this analysis interms the potential for distinguishing the relative contributionsof habitual and goaldirected processes to typing decisionsbetween individuals and at different levels of control.
Keywords: Errors, ActionSlips, Habits
JOHANNES RITTER University of Erfurt
A basic tenet of the bounded rationality approach holdsthat processing of decisionrelevant information is effortfuland that processing resources are limited. Building on thisassumption, several research strands have shown that peoplecan selectively use pieces of information in their decisionmaking – as opposed to exhaustive use of all information –and thereby save effort and time. In contrast, there areseveral experimental paradigms (e.g., hindsight bias, strooptask, etc.) suggesting that the suppression of readilyavailable information itself is effortful. In addition, there iswidespread evidence that the human cognitive system isable to integrate several pieces of information effortlesslyand automatically even across different modalities. Thus,there are two opposing hypothesis concerning theconsequence of cognitive exhaustion on decision making.The bounded rationality perspective predicts omission ofinformation and employment of parsimonious decisionstrategies such as heuristics. The automated integrationperspective predicts more exhaustive use of informationavailable in the decision making process.
The present study was designed to test these opposinghypotheses using a cognitive exhaustion paradigm:Participants in the exhaustion condition were required toperform a colorstroop task before and in between blocks ofprobabilistic decision tasks. Participants in the controlcondition were required to perform a nonexhaustive versionof the stroop task before and in between blocks ofprobabilistic decision tasks. Additionally to the exhaustionfactor, the decision environment was manipulated with a lowdispersion of cue validities and high dispersion of cuevalidities, resulting in a 2x2 design with the factors exhaustion(high vs. low) and decision environment (high dispersion vs.low dispersion). High dispersion environments are assumed tostipulate selective decision strategies such as noncompensatory heuristics because these environments allowomitting pieces of information without a decline in accuracy.It was therefore of special interest for the hypotheses howparticipants would decide in the condition with highdispersion and under high exhaustion. The decisions of theparticipants were analyzed using decision outcomes andconfidence judgments. Participants were classified as users ofeither an exhaustive decision strategy (using all information)or a selective decision strategy (using only limitedinformation). As expected, the results show that participantswere more likely to use selective decision strategies in thehighdispersion environment. However, they were less likely to
DMB 2014 | 912 Sept 2014 | 111
do so under cognitive exhaustion. These results are more inline with the automated integration perspective and call theassumption of generally limited cognitive resources of thebounded rationality perspective into question.
Keywords: bounded rationalty, cognitve exhaustion,probabilistic inference
HRVOJE STOJIC Universitat Pompeu FabraPANTELIS P. ANALYTIS Max Planck Institute for HumanDevelopmentMEHDI MOUSSAI Max Planck Institute for HumanDevelopment
From academic search engines like Google Scholar tonews sharing tools such as Digg many of our choices takeplace in environments that rank the alternatives in decreasingorder of popularity. The order in which the decision makersconsider the alternatives has been shown to be an importantelement of the decision making process. What happens whenindividual choices alter the popularity of the alternatives,and consequently, the order in which other decision makersconsider them?
In this paper, we present a simple collective behaviormodel in which the individual behavior is based on theprinciples of search theory. In our framework decision makerswith correlated preferences decide sequentially whichalternative to choose. The decision makers sample thealternatives in order of decreasing popularity and choosethe first satisficing alternative they encounter. The popularityorder is updated after each individual choice. In comparisonto existing social influence accounts of multialternativechoice our model is the first to explicitly describe the searchand stopping rules followed by the decision makers.
We present analytical results for some boundaryconditions that are tractable and simulation results for therest. We simulate a market consisting of hundred alternativesand thousand agents varying the satisficing thresholdemployed by the agents and the degree of heterogeneity ofpreferences in the population. Throughout the analysis, weemploy an optimal stopping model in which the decisionmakers sample the alternatives at random as a theoreticalyardstick.
We find that in popularity search the decisions of fewdecision makers in the beginning of the sequence determinethe search path for decision makers choosing after them andhave a large impact on the overall outcome in the market.This effectively leads to richgetricher dynamics, which implygreater unpredictability and inequality in the market ascompared to a random search environment, where the marketshares of the alternatives are predictable in the long run. Weshow that in popularity search a lower threshold implies moreinequality in the market while it has the exactly oppositeeffect in random search. Further we study the strength of therichgetricher dynamics by comparing the utility of the mostpopular alternatives in the market to that of the averagealternative utility and the highest possible utility. Finally, inpopularity search we find that some preferenceheterogeneity leads to higher levels of average utility as
compared to a population with perfectly homogenouspreferences. We explain this result by showing how someheterogeneity implies additional search, which in turn leadsto capturing some of the informational externalities.
Keywords: orderedsearch, satisficing, collective behaviour,richgetricher dynamics
JIEYU LV Queen Mary University of LondonMICHAEL J. PROULX University of BathMAGDA OSMAN Queen Mary University of London
“What can one do to promote cooperation?” As themost frequent prosocial behaviour, cooperation is a wellstudied topic in psychology, and has led many to take upthis question. Several methods have been used to promotecooperation, such as the inclusion of reinforcers (i.e. rewardsand punishments) (Fehr & Gächter, 2000). Others havefocused on methods designed to promote trust, identify withothers, and increase people’s perspectives of long termprospects (Parks et al., 2013). Along with these factors, onewhich has been argued to be crucial for supportingcooperation is empathy (Batson & Ahmad, 2001; Batson etal., 1995; Batson & Moran, 1999; Beadle et al.,2013; Cohen& Insko, 2008; Rumble et al., 2010; Xu et al., 2012). In orderto address how empathy and status effects cooperation, weused a classic social dilemmas paradigm Public GoodsGame (PGG).
The present study included two computerbased PGGexperiments. There were three fictitious confederates andone real participant involved. Experiment 1 (N= 52): Weused a betweensubject design with three levels of empathy(highempathy [HE, n = 17], lowempathy [LE, n = 17] and noempathy [NE, n=18]) and randomly allocated to eachcondition. The empathy manipulation involved breaking upwith boyfriend, experiencing a car accident, and having amobile phone stolen; this was similar to Batson and Moran’s(1999) empathy induction method. After this, there was aPGG, and finally a set of empathy questionnaires werepresented (Interpersonal Reactivity Index (IRI) (Davis, 1983).We found that the empathy manipulations didn’t yieldsignificant differences in the average contributions in thePGG [F (2, 49) = 0.101, p = 0.904 > 0.05)]. Post hoc testshowed that there were no significance differences amongthe three conditions; HE (M = 9.00, SD = 7.29), LE condition(M = 9.08, SD = 7.29), NE condition (M = 8.50, SD = 7.08).
In Experiment 2, just as in Experiment 1 there were threefictitious confederates and one real participant wereinvolved (N = 49). We compared highstatus [n = 26] withlowstatus [n = 23] conditions. There was no manipulation ofempathy. Participants simply saw pictures and names of threefictitious characters. All three fictitious confederates wereendowed with 20 tokens at the start of the PGG, but half ofthe real participants were endowed with 30 tokens (highstatus) and the other half were endowed with 10 tokens (lowstatus). We found that consistent with our prediction, the lowstatus group (M = 0.565, SD = 0.28) contributed morepercentage than the high status group (M = 0.441, SD =0.301) [F (1, 47), p = 0.041 < 0.05].
112 | DMB 2014 | 912 Sept 2014
In summary, in Experiment 1 we found that the empathymanipulation didn’t increase cooperative behaviours in thePGG. A further replication would be needed to fully establishhow reliable this finding is, and whether empathymanipulations in general are an effective way of inducinghigher level of cooperation. Batson and Moran (1999) andRumble et al. (2010) have shown that empathy inducescooperation, but other work suggests that the role ofempathy in cooperation is less reliable (Eisenberg & Strayer,1990, p. 301). In Experiment 2, we found individuals in thehigh status position behaved less cooperatively than thosein the low status position. Future work could explore potentialinteractions between empathy and status. It may well be thatthe empathy induction methodology that we used requiresfurther finessing, and establishing a link between empathyand status as indexed by levels of cooperation would be aninnovation in the domain of research on prosocialcooperative behaviours.
Keywords: empathy; status; cooperation; Public GoodsGame
FERYAL M. P. BEHBAHANI Imperial College LondonA. ALDO FAISAL Imperial College London
Intelligent behaviour is fundamentally tied to the ability ofthe brain to make decisions in uncertain and dynamicenvironments. In order to accomplish this task successfully, thebrain needs to categorise novel stimuli in realtime. Inneuroscience, the generative framework of Bayesian DecisionTheory has emerged as a principled way to predict how thebrain has to act in the face of uncertainty (Ernst & Banks,2002, Kording & Wolpert, 2004, Kemp & Tenenbaum, 2008,Faisal et al., 2008). We hypothesise that the brain might alsouse generative Bayesian principles to implement itscategorisation strategy. Previous experimental work on humancategorisation shows data that is consistent with bothdiscriminative and generative classification (Hsu & Griffiths,2010) and did not allow confirming the implementation ofone or the other. Therefore, we designed a novel experimentin which subjects are trained to distinguish two classes A andB of visual objects, while exemplars of each class are drawnfrom Gaussian parameter distributions, with equal varianceand different means. During two different experimentalparadigms, we test how the subjects representation of thecategories change as a result of being exposed to outliersfor only one of the categories, A, far from category B, i.e.,increasing category A's variance.
Generative classifiers are by necessity sensitive to novelinformation becoming available during training, whichupdates beliefs regarding the generating distribution ofeach class. In contrast, discriminative classifiers are sensitiveto novel information only if it affects the immediatediscrimination of classes. In the first paradigm, wecharacterise the categorisation boundary, i.e., the pointwhere a subject assigns with equal probability a test stimulusto either category. Subsequently, we track the shift in theboundary after the introduction of outliers. A generative
classifier will prompt to reconsider the variance of class Aafter being exposed to the outliers and accordingly shifts thecategorisation boundary towards category B. However, thediscriminative classifier will not react, as there is no newinformation added to the boundary itself. Our secondparadigm provides an even more stringent test forgenerative models: again, outliers for class A are presentedfar away from class B. Additionally, the two classes areselected to be close enough, such that a generativeclassifier would assume that class A's variance has increasedsignificantly, thereby reaching across the region occupied byB. This will result in the emergence of a second classificationboundary to the distal side of class B, far away from class A.Again, the discriminative classifier would not change itsbehaviour.
Our results in both paradigms show that the introductionof the outliers for category A influences the subject'sknowledge of the distribution associated with alternativecategories. This can result in the introduction of additionalboundaries only predicted by our simulations of generativeclassifiers. These results give clear evidence that visualcategorisation is only consistent with generative and notdiscriminative classification mechanisms. Furthermore, ourexperiments provide an ideal experimental framework forneurophysiological and functional imaging investigations ofthe underlying neural mechanisms involved.
Keywords: categorisation, visual perception, generativemodels
RICHARD FAIRCHILD University of Bath
Behavioural economists are increasingly understandingthat humans are not completely selfinterested or emotionless,but often exhibit “otherregarding” behaviour. We develop agametheoretic approach in which players gain utility fromtheir own material payoffs, but who also develop empatheticemotions towards each other. David Sally (2001, 2002)argues that reciprocal behaviour may depend on the socialcontext, social interaction, and psychological closeness ofthe players. Motivated by Sally ’s seminal analysis of sympathygames, we develop a twostage, extended form, empathymodel in which players simultaneously choose empathy levelsin one stage, and, in a separate stage, make simultaneousstrategy choices in a material game. We consider bothconscious (strategic/instrumental), and unconscious (innate)empathy (you simply like someone for who they are). Wedemonstrate that increasing empathy can have positive ornegative effects on welfare in equilibrium, and that theseequilibria can be crucially affected by whether empathy isformed consciously, or unconsciously. We tentatively term ourmodelling approach, ‘emotional game theory.’
Keywords: Game theory; empathy; conscious andunconscious emotions
DMB 2014 | 912 Sept 2014 | 113
NEIL STEWART University of WarwickTAKAO NOGUCHI University of WarwickCHRISTOPHER Y. OLIVOLA Carnegie Mellon UniversityHELEN SUSANNAH MOAT Warwick Business SchoolTOBIAS PREIS Warwick Business School
People's Internet searching reveals something of theirthinking. Search engines such as Google offer access todaily aggregated frequencies for search terms at the level ofnations and states. We used Google Trends to extract (a)psychological measures of people's thinking about time and(b) measures of the connotation of people's searches. Thesemeasures correlate strongly with the economic performanceof nations and states, demonstrating the viability ofconstructing nationlevel psychological measures fromGoogle Trends data.
Time Perspective: from the frequencies of 4digit yearsearch terms (e.g., "2012") we constructed fourpsychologically inspired measures of timeperspective andexamined their relationship with a widelyused measure ofeconomic activity, percapita gross domestic product (GDP).Past and future focus measures the extent to which peoplesearch for last year or next year compared to this year. Atthe level of individuals, higher future focus is be associatedwith stronger economic performance (Zimbardo & Boyd,1999). Past and future time horizon measures how far into thepast or future people are searching. At the level ofindividuals, thinking further into the future is associated withbeing richer and more educated (for review, see Reimers,Maylor, Stewart, & Chater, 2009). We find that nations withhigher percapita GDP are more focused on the future, χ2(1)= 19.65, p < .001, and less on the past, χ2(1) = 9.19, p =.002, and that when these nations do focus on the past, it ismore likely to be the distant past, χ2(1) = 7.77, p = .005(Figure 1). Our measures of timeperspective capture 53% ofthe variance in percapita GDP.
Connotations of Search Terms: the semantic differential(Osgood, 1952) is to concepts as the big five is topersonality. Using ratings of how well adjectives describeconcepts, Osgood recovered a threefactor solution for themeasurement of meaning. We used the relative frequencies ofOsgood's adjectives in search terms to construct a semanticdifferential for Google searches at the level of US states. Thesecond factor (kind, nice, sharp, deep, sweet vs. crule, awful,dull, shallow, bitter) showed a strong correlation with percapita gross state product (GSP) and equality of incomedistribution (Gini coefficient). More kind, nice, sharp, deep,and sweet searching is associated with lower GSP (r = .21, p= .0002, Figure 2) and a more equal distribution of income (r= .39, p = .004, Figure 3).
Although the direction of causality cannot beestablished from our correlational analyses, the strongassociations of our measures with known economic measureslike GDP and inequality indicate the viability of using nationlevel data, together with psychological concepts, tomeasure the behaviour of nations.
Keywords: Google trends, temporal discounting, grossdomestic product
ANJALI MAZUMDER University of WarwickJULIA BRETTSCHNEIDER University of Warwick
The rich heterogeneity and success of academicinstitutions can in part be reflected by universities ’ admissionsdecisionmaking process. Admissions tutors are tasked withmaking decisions regarding placement offers to applicantswith the expectation that an applicant will succeed in theprogramme of study. While institutions provide informationabout the applicant profile they seek, the actual decisionmaking process is often seen as less transparent. Admissionsdecisionmaking can be a complex process, often madedifficult by multiple and partial sources of information which isnot uniformly collected or used. A simplistic generalisation ofadmissions can be broadly described as (i) formulaic, relyingprimarily on academic results, or (ii) ‘holistic’, examiningbroader student profiles such as statements and references.Often decisions involve a more comprehensive approach inwhich formulaic criteria are combined with holisticjudgements: The challenge then becomes assigning ratingsto attributes of judgements to facilitate a structured,consistent and transparent approach.
A key challenge to making an offer is that decisions areoften made on predicted grades as at the time of applying,most students will not have final grades. This introduces a newdecisionmaking problem: Once an offer has been made,some applicants may fail to meet the threshold andsubsequently a decision to rescind the offer may be made.Other supplementary results can support an offer such asSTEP results (for the mathematical sciences). An addeduncertainty is that there may be variability in what a gradelevel achieved at one school is versus another. Holisticevidence such as participation in competitions (e.g. MathOlympiads), supporting statements and references, isinherently variable. Thus, decisionmaking involves a mixture ofmeasured variables and expert judgement.
While some institutions have longestablished practices,entrenched in historic culture and underlying missions,admissions has become increasingly dynamic. External factorssuch as changing demographics, economics, and political,social and educational environment have prompted manyinstitutions to modify their selection process. Shifts in earlydecisions, open days, interviews, etc. and technology e.g.,online applications, have encouraged institutions to rethinkhow to manage applications and effect admissions process.Students ’ decisions to apply and accept an offer is multifaceted: This may involve financial incentives, careerprospects, location, etc. Thus, therein lies the challenge ofachieving a careful balance of set targets and numbers(whether financial and/or demographic) and attractingapplicants who will succeed and fit into the institutionalenvironment.
We extract and use historically relevant data fromapplicants to mathematical sciences programmes atdepartmental, university and national level to inform a graphmodular structure for an admissions decision support system.Using Bayesian Networks, we develop a decisionsupportsystem to deal with uncertain, ambiguous and incoherentevidence. The graph modular structure facilitatesinterrogation of submodels, calibrating particular facets orsubobjectives. Furthermore, it aids and provides a vehicle for
114 | DMB 2014 | 912 Sept 2014
communicating the decisionmaking process within a team,ensuring a dynamic, systematic and transparent approach ismanaged over a long application period.
Keywords: decision support systems, uncertainty, expertjudgement, admissions, combining evidence
JASMIN MAHMOODI University of EdinburghELIZABETH J. AUSTIN University of Nottingham
A common decision making paradigm is delaydiscounting, which is a way of assessing financial decisionsregarding receiving either a smaller immediate or a largerdelayed monetary reward. Individuals differ in the amount ofbenefit required to cause them to choose the delayedoptions. Currently, the main candidates for accounting forthese differences are mood and the personality traitsimpulsivity and selfcontrol. To date, however, no study yethas examined the potential role of trait EmotionalIntelligence (EI), which can be described as the ability torecognise, modify, and use own and others ’ emotions, indelay discounting. This study investigated the relationship ofEI to individual differences in delay discounting, contrastingthis with possible effects of induced mood. Contrary toprevious findings, neither positive nor negative inducedmood influenced discounting behaviour. Turning to theeffects of EI, regression analyses showed that the “Regulationof Emotion” facet of trait EI significantly predicted lowerdiscount rates, which is the preference for larger delayedrewards. No similar effects were found for either total EI norother EI subscales. Results are discussed against thebackground of previous research and its limitations andimplications for future research are addressed.
Keywords: delay discounting, emotional intelligence, mood,decision making, individual differences
ROBERT BIEGLER NTNUGERIT PFUHL NTNU
Making good decisions depends on rationality as muchas or even more than on intelligence. Stanovich (2009)pointed out that if rationality is lacking, intelligence may onlyprovide people with more plausible reasons for believingwhatever it is they want to believe. When it comes toselecting people for various positions, intelligence is oftentested, but rationality at best rarely, perhaps never. Onereason for that state of affairs is that intelligence tests arewell developed, reliable, and familiar.
We have found only one rationalitytest in the literature(Pacini and Epstein, 1999). It consists of 40 items thatrecord attitudes towards and selfassessments of rationality.The test does not directly measure rationality itself.
Scattered across a number of papers are items that seem tomeasure rationality more directly.
We want to see how well these correlate with Pacini andEpstein’s test, and also with measures of simple forms ofmetacognition: how good are people at judging how wellthey remember, and how good are they at estimatingambiguity?
We measured a) how well people can judge how precisetheir memory for locations and shapes is; b) their ability toestimate conditional probabilities given varying numbers ofdata points; c) how they perform on the rationalityquestionnaire items e.g. knights and knaves problem. Wefurther examined how much evidence people need to drawa conclusion, and how much they dislike uncertainty in theforms of risk and of ambiguity.
Finally, we correlated the results with an IQ test. We founda significant correlation between IQ and rationalengagement and Stanovich’s rationality measure. Further,rational engagement was correlated with Need for Closure.There was no effect of attentional control or risk andambiguity.
Keywords: metacognition, cognitive biases, uncertainty, risk,ambiguity, IQ
CASIMIR LUDWIG University of BristolDAVID R EVENS University of Bristol
How do we sample noisy information from multiple sourcesin order to make decisions? We address this general problemin the perceptual domain. Subjects make a comparativedirection discrimination judgement of two random dot motionpatterns that are activated only when directly fixated. Theviewing time is limited and subjects have to adapt their timeallocation to the quality of the two patterns. A simpleforaging model assumes a pattern is sampled until thechange in the precision of the direction estimate for thatpattern drops below a threshold. At that point the subjectswitches to the other pattern and accrues information fromthat source. We test two predictions of the model: (i) thetiming of the first switch should depend on the online accrualof information; (ii) the representation of previously sampledinformation decays.
We tested the first prediction by briefly “pulsing” thequality of motion information during the fixation of the firstpattern. If the switch time depends on the amount ofinformation accrued (i.e. amount of certainty about thedirection of the first pattern), subjects should switch earlierwhen the motion coherence of the first pattern is brieflyenhanced. Conversely, when the information quality is brieflydegraded, subjects should delay the timing of the first switch.
We tested the second prediction by asking subjects toestimate the direction of either the first pattern or the secondpattern. Subjects are directed to inspect one pattern firstand when that pattern stops moving, they have to switch tothe second pattern. We varied the viewing duration of boththe first and the second patterns. Once both patterns havebeen fixated, a dial appears in the location of either the firstof the second pattern. Subjects have to estimate thedirection of the corresponding pattern. Decay is estimated
DMB 2014 | 912 Sept 2014 | 115
through the precision of the direction estimate of the firstpattern, as a function of the viewing duration of the secondpattern.
Experiment 1 confirms that in the comparative decisiontask, switch time is (at least partly) controlled by the onlineaccrual of information. Experiment 2 demonstrates that thereis surprisingly little decay in the representation of previouslysampled information. However, the accrual of new informationis slowed down as a result of having to maintain arepresentation of the previously sampled source. Thesefindings place important constraints on our model ofinformation foraging.
ANDREW HIGGINSON University of BristolDANIEL NETTLE University of BristolBEN BRILOT University of Bristol
Anxiety disorder is usually assumed to be a product of amalfunctioning cognitive system. Adaptive reasons that acognitive system may be disposed to anxiety disorder haverarely been considered. We modelled a Bayesian foragerthat learns about the presence and density of predators inits environment. The forager decides the proportion of its timeto look for predators rather than food and what level of falsealarms to tolerate. This optimal learning system results in asignificant proportion of individuals behaving as if theenvironment is dangerous even when predators have beenrare for a long time, which we consider to be operationallyanalogous to anxiety disorder in humans.
Keywords: Predation, Foraging, Bayesian learning, Signaldetection
DMB 2014 | 912 Sept 2014 | 116
DMB 2014 | 912 Sept 2014 | 117