THE PRELIMINARY ORAL EXAMINATION FOR THE DEGREE OF Ph.D. IN COMPUTER SCIENCE FOR

                           James Ryan Carr

Evolutionary game theory has frequently been used to investigate behavioral phenomena   observed in nature. Cultaptation is an evolutionary simulation game that aims to investigate the phenomenon of social learning, or learning from the behavior of others. Games like Cultaptation have proven significantly more difficult to analyze than classical EGT problems, due to the complications introduced by allowing social learning: agents can live for multiple generations, interact with one another, and their fitness at each generation is determined by all of their previous actions.

My work has had two main objectives: (1) to study the nature of Cultaptation to see what types of strategies are effective; and (2) more generally, to develop ways of analyzing evolutionary environments with social learning. My results to date include the following:

1. I show how lookahead search techniques can be used to analyze the reproductive success of existing strategies by calculating their expected utility to within any epsilon > 0.

2. I show how to compute a near-best response strategy for Cultaptation using a finite-horizon lookahead search.

3. I present the Cultaptation Strategy Learning Algorithm, which attempts to find a Cultaptation strategy that is a near-best response to itself, and is therefore a symmetric near-Nash equilibrium.

4. I present experiments that examine the nature of strategies generated with CSLA, and show how such experiments can be used to glean insights into the nature of Cultaptation and social learning in general.

My proposed extensions to this work include:

1. I will develop an improved version of CSLA, which will use analytic methods in place of the simulations used by the current version. This will make it possible to prove that either CSLA returns an epsilon-Nash equilibrium strategy or no such strategy exists for the given parameters, which will provide a much stronger justification for using strategies generated by CSLA to examine properties of social learning in Cultaptation.

2. I will use the insights gleaned from my experiments with CSLA strategies to design a strategy to play in the Cultaptation tournament. This strategy will perform a relaxed lookahead search that will allow it to approximate the effects its actions will have several rounds in the future, which will give it an advantage over the greedy, discount-reward scheme used by the current best strategies. If successful, this strategy will exhibit how the lessons learned from examining CSLA strategies can be applied to improve the social learning capabilities of agents in other environments

Examining Committee:

Dr. Dana Nau                                         -          Chair

Dr. V.S. Subrahmanian                          -          Dept’s Representative

Dr. Michele Gelfand                              -          Committee Member

EVERYBODY IS INVITED TO ATTEND THE PRESENTATION