Game Theory, Social Interactions and Artificial Intelligence Supervisor: Philip Sterne Supervisee: John Richter
Contents ● Quick and Dirty Summary ● Outline of Game Theory ● Outline of Algorithms used ● Discussion of Feasability ● Implementation Details ● Readings and Sources of Information ● Proposed Timeline ● Questions
Quick and Dirty Summary Using a variety of artificial intelligence techniques to develop optimal strategies for participants to use in situations where some combination of competition and cooperation are necessary.
Game Theory ● Prisoner's Dilemma ● Simple – Only two participants – Only two choices ● Because of simplicity, it is trivial to consider all possible eventualities “The Evolution of Cooperation”, Robert Axelrod, Basic Books, 1984
Artificial Intelligence ● Emulating features of human intellect ● Particularly useful for solving non-trivial problems – function maximization, optimization ● Contemporary AI forms: – Expert systems, Case-based reasoning – Neural Networks – Genetic Algorithms – Reinforcement Learning
AI: Artificial Neural Networks ● Modelling the actual structure of the human brain ● The neurons making up the map use weights and thresholds to emulate a complex function ● The map itself is quite complex to create, with neuron layers of various sizes connected together ● Of minor use in my project
AI: Genetic Algorithms ● Used to maximize functions ● “Breeds” the best solutions ● Uses crossover (joining solutions at random points), mutation (occasionally changing a random value to something else to ensure global maximum found) on a variety of sub- optimal solutions to encourage the best of them to breed, and results in the next generation representing better solutions
AI: Reinforcement Learning ● Most recent of these AI's ● Similar to GA's – approaches best solution ● Unlike GA's – learning happens during the agents life, not (always) passed on to the children ● Uses reward/punishment system to encourage agents to take best strategies, while still encouraging discovery and innovation “Reinforcement Learning, an Introduction”, Sutton and Barto, MIT Press, 1998
Less Quick, Less Dirty Summary Using genetic algorithms and reinforcement learning to train neural nets in agents in order to model complex game theory, and developing optimal strategies for real-world games
Feasability ● Is the project too small? – Can be extended by testing other AI approaches and combinations – Can be used to model a variety of scenarios ● Is the project too big? – Three distinct portions can be simplified: ANN's, GA's and RL. – Simplify the test model to use less variables, less complex ANN's
Implementation ● No specific languages or platforms are required or recommended for developing AI ● Thus, I choose the route of free software and mainstream languages, and shall write the program for the GNU/Linux platform, in C++
Further Resources and Reading ● JASSS, Journal of Artificial Societies and Social Simulation [ ● “Design, Evaluation and Comparison of Evolution and Reinforcement Learning Models”, Clinton Brett McLean (2001) ● Game Theory.Net [ ● Hours and hours of playing games in the labs. No, really.
Timeline
Questions ● To maintain order, could I request questions in phases: – AI algorithms or techniques – Game theory – Integration – Feasability and implementation of project