Crowdsourcing NP-Complete Problems on the Web (Presentation Slides)

2
Swinburne University of Technology 1 Crowdsourcing NP-Complete Problems on the Web!? Jason Brownlee, Daniel Angus Complex Intelligent Systems Human-Driven Computational Intelligence? Last Monday: What if… Humans solving localised problems (like Ants? Birds? etc..) Use the Web as the Interface Aggregate contributions into a holistic solutions Could we solve really hard problems? NP-Complete? AI-Complete? Why the Web? Web: Lots of humans with lots of time Examples: Mechanical Turk ($ for click labour) Citizen Science galaxyzoo, … Games with a Purpose reCAPTCHA, ESP, Peekaboom, etc. Problem & Approach TSP: T ravelling S alesman P roblem NP-Complete, Familiar, Well Understood Easily Partition into Sub-Problems (sub-tours) General Approach Do Something like A nt C olony O ptimisation with Humans Work on Sub-Problems: Localised Stepwise Construction Exploit human pattern recognition!!! (underlying structure?) Top TSP Picture from Wikipedia: http://en.wikipedia.org/wiki/Trav elling_salesman_problem From To TSP System Overview Database (1) Sub-Problem Selection (2) Solve Sub-Problem (3) Aggregation (Adjacency Matrix) (4) Holistic Solution Random, Spatial, etc… Connect-the-dots, play a game? Greedy, Probabilistic, etc… Is The System Viable? Is a human at least as good as Nearest Neighbour? I Suspect Better: Innate Spatial Intelligence! If so, does the system work under this assumption? What Sub-Problem Selection Scheme? What if Human Users Suck? What is happening inside the database (adjacency matrix)?

description

These are the presentation slides used by Jason Brownlee to present the humanTSPsolver project to the complex intelligent systems lab, Swinburne University in 2008. The slides review the premise of the project which is that a crowdsource model may be useful in solving very complex problem instances.

Transcript of Crowdsourcing NP-Complete Problems on the Web (Presentation Slides)

Page 1: Crowdsourcing NP-Complete Problems on the Web (Presentation Slides)

Swinburne University of Technology 1

Crowdsourcing NP-Complete Problems on the Web!?

Jason Brownlee, Daniel AngusComplex Intelligent Systems

Human-Driven Computational Intelligence?� Last Monday: What if…

� Humans solving localised problems (like Ants? Birds? etc..)� Use the Web as the Interface� Aggregate contributions into a holistic solutions

� Could we solve really hard problems?� NP-Complete? AI-Complete?

Why the Web?� Web: Lots of humans with lots of time� Examples:

� Mechanical Turk ($ for click labour)� Citizen Science

� galaxyzoo, …

� Games with a Purpose� reCAPTCHA, ESP, Peekaboom, etc.

Problem & Approach� TSP: Travelling Salesman Problem

� NP-Complete, Familiar, Well Understood� Easily Partition into Sub-Problems (sub-tours)

� General Approach� Do Something like Ant Colony Optimisation with Humans� Work on Sub-Problems: Localised Stepwise Construction� Exploit human pattern recognition!!! (underlying structure?)

Top TSP Picture from Wikipedia: http://en.wikipedia.org/wiki/Trav elling_salesman_problem

From To

TSP System Overview

Database

(1) Sub-Problem Selection

(2) Solve Sub-Problem

(3) Aggregation(Adjacency Matrix)

(4) Holistic Solution

Random, Spatial, etc…

Connect-the-dots, play a game? Greedy, Probabilistic, etc…

Is The System Viable?� Is a human at least as good as Nearest Neighbour?

� I Suspect Better: Innate Spatial Intelligence!

� If so, does the system work under this assumption?� What Sub-Problem Selection Scheme?� What if Human Users Suck?� What is happening inside the database (adjacency matrix)?

Page 2: Crowdsourcing NP-Complete Problems on the Web (Presentation Slides)

Swinburne University of Technology 2

What Sub-Problem Selection Scheme?

Spatial sub-problem generation is generally better than Random!

What if Human Users Suck?

Garbage-in, Garbage-out, as expected.

What is Happening Inside the Database?

Collecting interesting information, how do we use it effectively?

Where the Project is at:� Java Codebase (Experimentation with OAT)� Web Application Prototype (Ruby on Rails)� 4 Technical Reports (with 3 Experiments)

Open Problems: Any Thoughts?Collect Useful Information

� Visually Motivate With Heuristics? History? How Much?

Extract Useful Information� Emergent Solution from Degenerate Sub-Solutions?

Motivate Participation� Drawing? Game? Competition? Gimmick?� Fly a Spaceship? Topography? Polynomial Transformations?