AI for Social Good:

Key Techniques, Applications, and Results

MILIND TAMBE

Founding Co-director, Center for Artificial Intelligence in Society (CAIS)

University of Southern California

tambe@usc.edu

Co-Founder, Avata Intelligence

USC Center for Artificial Intelligence in Society (CAIS.USC.EDU)

2/11/2017 2

Mission Statement: Advancing AI research driven by…

Grand Challenges of Social Work

Ensure healthy development for all youth

Close the health gap

Stop family violence

Advance long and productive lives

Eradicate social isolation

End homelessness

…

2/11/2017 3

Overview of CAIS Project Areas

AI for Assisting Low Resource Communities

Social networks: Spread HIV information

Maximize influence under uncertainty

Real-world pilot tests: Big improvements

AI for Protecting Endangered Wildlife

Machine learning/planning: Anti poaching

Scale, boundedly rational poachers,…

Real-world: Uganda, South Asia…

2/11/2017 4

Gangs, Substance abuse,

Veterans mental health

Social networks, robust optimization,…

Behavioral models…

Real-world: Research in progress

AI for Public Safety and Security

Game theory: security optimization

Solve massive “security games”

Real-world: US Coast Guard, FAMS…

Overview of CAIS Project Areas

2/11/2017 5

Outline

Introduction

HIV Information among homeless youth

Wildlife Conservation

2/11/2017 6

AI Program: HEALER

2/11/2017 7

Outline: HIV Information & Homeless Youth

Domain of homeless youth and HIV information dissemination

Real World Challenges in Influence Maximization

POMDP Model and algorithms

Pilot Study

2/11/2017 8

Influence Maximization Background

Input:

Graph G

Influence Model I

Choose K nodes per time step

Number of time steps for influence spread T

Output:

K nodes per time step maximizing expected # influenced

nodes

2/11/2017 9

Independent Cascade Model

Propagation Probability (for each edge)

2/11/2017 10

Real World Challenges

Uncertain network state

Uncertainty in network structure

Adaptive selection

2/11/2017 11

Challenge 1: Uncertain Network State

B

C

2/11/2017 12

Challenge 2: Uncertain Network Structure

2/11/2017 13

Independent Cascade Model

Propagation Probability (for each edge)

Existence Probability (for uncertain edges only)

2/11/2017 14

HIV Prevention Programs:Using Social Networks to Spread HIV Information

2/11/2017 15

Challenge: Adaptive selection in Uncertain Network

K = 5

1st time step

2/11/2017 16

Challenge: Adaptive selection in Uncertain Network

K = 5

2nd time step

2/11/2017 17

Challenge 3 : Adaptive selection

K = 5

3rd time step

2/11/2017 18

NO LONGER A SINGLE SHOT

DECISION PROBLEM

Outline: HIV Information & Homeless Youth

Domain of homeless youth and HIV information dissemination

Real World Challenges in Influence Maximization

POMDP Model and algorithms

Pilot Study

2/11/2017 19

POMDP Model

Sequential decision making under uncertainty

Homeless shelters – sequentially select nodes

Homeless shelters – network state not known

Action

Choose nodes

Observation: Which edges exist?

Adaptive Policy

REWARD

POMDP SOLVER

HIDDEN

World State: Actual

node/edge state

2/11/2017 20

Real world scale: Why is it hard to solve?

2300 states150C6 actions

Current offline and online

POMDP solvers unable to

scale

2/11/2017 21

Real world scale: Why is it hard to solve?

2/11/2017 22

Real world networks have community structure

Graph Partitioning

2/11/2017 23

Graph Partitioning

2/11/2017 24

HEALER : Hierarchical Ensembling

GRAPH PARTITION TOOL

. . . .

. . . .

GRAPH SAMPLER

ORIGINAL POMDP

Intermediate

POMDP

Sampled

POMDP

Sampled

POMDP

Sampled

POMDP

HEALER

Graph

SamplingGraph

Sampling

Graph

Sampling

Intermediate

POMDP

Intermediate

POMDP

2/11/2017 25

Graph Partitioning: Why partition?

qINTERMEDIATE

POMDP

2/11/2017 26

Real Networks - Solution Quality

2/11/2017 27

Robustness & Parameter Uncertainty

HEALER: fixed propagation and existence probability

Want policies robust to different possible values of P(A,B) and U(A,B)

Express as ranges of values, e.g., U(A,B) is in [0.4, 0.8]

2/11/2017 28

HEALER++ Robustness & Parameter Uncertainty

Worst case parameters: a zero-sum game against nature

Payoffs: (performance of algorithm)/OPT

Nature

Chooses parameters

P(A,B) and U(A,B)

vs

Algorithm

Chooses policy in

POMDP state-action

space

2/11/2017 29

Outline: HIV Information & Homeless Youth

Domain of homeless youth and HIV information dissemination

Real World Challenges in Influence Maximization

POMDP Model and algorithms

Pilot Study

2/11/2017 30

Pilot Tests with 170 Youth in LA Area

Recruited youths:

Preliminary network —> HEALER

Bring 4 youth for training, get edge data —> HEALER

Bring 4 youth for training, get edge data —> HEALER

Bring 4 youth for training

HEALER HEALER++ DEGREE CENTRALITY

62 56 55

2/11/2017 31

Results: Pilot Studies

0

20

40

60

80

100

HEALER HEALER++ Degree

Percent of non-Peer Leaders

Informed Not Informed

2/11/2017 32

Results: Pilot Studies

0

20

40

60

80

100

HEALER HEALER++ Degree

Informed Non-Peer Leaders Who Started Testing for HIV

Testing Non-Testing

2/11/2017 33

AI Program: HEALER

2/11/2017 34

Next Steps

900 youth study begun at three locations in Los Angeles

300 enrolled in HEALER/HEALER++

300 enrolled in no condition

300 in Degree centrality

IRB approvals

Presenting video updates every few months

2/11/2017 35

Outline

Introduction

HIV Information among homeless youth

Wildlife Conservation

2/11/2017 36

Uganda Wildlife Authority & Wildlife Conservation Society

Protecting Wildlife in Uganda

2/11/2017 37

Predicting Poaching from Past Crime Data

PAWS: Applying AI for protecting wildlife

Poacher Behavior Prediction

2/11/2017 38

Queen Elizabeth National Park, Uganda

Poacher behavior prediction

1300 targets, 12 years of patrol data

How likely is an

attack on

a grid Square

Ranger patrol

frequency

Animal density

Distance to

rivers / roads

Area habitat

Area slope

…

2/11/2017 39

Initial Attempt Using Dynamic Bayes Net:Time Dependency & Imperfect Observation

Attacking probability

Detection probability

Ranger observation

Ranger patrol

Animal density

Distance to rivers /

roads / villages

Area habitat

Area slope

…

M

j

jxSEU

jxSEU

j adversary

adversary

e

eq

1'

)',(

),(Adversary’s

probability of

choosing target j

Interpretability, Speed

2/11/2017 40

Poacher Behavior Prediction

Poacher behavior prediction

Classifier 1 Classifier 2 Classifier 3

0 1 1

Aggregation Rule

1

Majority

1

Ensemble of Decision Trees

2/11/2017 41

0

0.5

1

1.5

2

2.5

3

3.5

4

L&L

Uniform Random SVM CAPTURE Decision Tree BoostIT-5Experts-Maj

Poacher Behavior Prediction

Poacher attack prediction

Results from 2015

2/11/2017 42

Real-world Deployment: Results

Two 3 sq km patrol areas: Predicted hot spots with infrequent patrols

Trespassing: 19 signs of litter etc

Snaring: 1 active snare

Poached Animals: Poached elephant

Snaring: 1 elephant snare roll

Snaring: 10 Antelope snares

Hit rates (per month)

Ours outperforms 91% of months

Historical Base Hit

RateOur Hit Rate

Average: 0.73 3

2/11/2017 43

2/11/2017 44

On-Going Experiments:Queen Elizabeth National Park

• Red: Group 1 (highest attack prediction rate)

• Yellow: Group 2

• Green: Group 3

2/11/2017 45

On-Going Experiments:Queen Elizabeth National Park

• Rangers followed poachers’ trail; ambushed camp

Arrested one (of 7) poachers

Confiscated 10 wire snares, cooking pot, hippo

meat, timber harvesting tools.

• Indirect poaching signs; pursuit of poachers

• Signs of road building, fires, illegal fishing

2/11/2017 46

Learn from crime data

Game Theory calculate

randomized patrols

Patrollers execute patrols

Poachers attack targets

Predicting Poaching from Past Crime Data

PAWS: Applying AI for protecting wildlife

Game Theory + Poacher Behavior Prediction

Towards the Future

Significant potential: AI for low resource communities, emerging markets

Direction of AI research in our hands

Not just applications; novel research challenges:

Fundamental computational challenges from use-inspired research

Designing AI systems in society:

• Interpretability

• Maintaining human autonomy

Methodological challenges:

Encourage interdisciplinary research: measures impact in society

2/11/2017 47

AI for Social Good

2/11/2017 48

THANK YOU

BACKUP

POMDP Model: Challenges

Number of states (node & edge uncertainty): ~ 2300

Number of actions (N-choose-K): > 1,000,000,000

Number of observations (Edges exist or not): Exponential

2/11/2017 51

Action

Choose nodes

Observation: Which edges exist?

Adaptive Policy

REWARD

POMDP SOLVER

HIDDEN

World State: Actual

node/edge state

Mission Statement: Advancing AI research driven by…

2/11/2017 52

Divide and Conquer

Main idea:

Divide POMDP into smaller POMDPs

Combine solutions of smaller POMDPs

Two different ways of dividing POMDPs

Using uncertain edges: PSINET

Using graph partitioning and sampling: HEALER

2/11/2017 53

…And the Past

“…prize every invention of

science made for the

benefit of all”

2/11/2017 54

Intermediate POMDP

Sampled

POMDP

Sampled

POMDP

Sampled

POMDP

.

.

2/11/2017 55

Intermediate POMDP

Sampled

POMDP

Sampled

POMDP

Sampled

POMDP

.

.

2/11/2017 56

Params #1 Params #2

Policy #1 0.8, -0.8 0.3, -0.3

Policy #2 0.7, -0.7 0.5, -0.5

HEALER++ Algorithm

Computes an equilibrium strategy for this game

Exponentially large strategy space: incremental generation, double oracle

Under some conditions, provably converges to approx. equilibrium

Params #1 Params #2

Policy #1 0.8, -0.8 0.3, -0.3

Policy #2 0.7, -0.7 0.5, -0.5

Nature’s oracle

Params #1 Params #2 Params #3

Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4

Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6

Influencer’s oracle

Params #1 Params #2 Params #3

Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4

Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6

Policy #3 0.6, -0.6 0.4, -0.4 0.7, -0.7

2/11/2017 57

Real networks - robustness

50

70

90

Venice Hollywood

Worst case % of optimal influence

HEALER++ HEALER

2/11/2017 58