The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9914

Title:Reinforcement Learning within a Multi-Agent Adversarial Soccer
Date: 1999
Abstract:The use of reinforcement learning (RL) for agent learning of sequential decision tasks has proved successful in many domains. In a series of experiments, Schmidhuber et al. applied several RL algorithms to a multi-agent adversarial soccer domain with the aim of learning successful multi-agent cooperative strategies. In this thesis, I successfully replicated Schmidhuber et al.'s work in applying the CMAC with World Models algorithm to a 3 vs. 3 soccer task. An exploration of the replicated task and learned policies showed that due to poor environmental design the learned policies are simplistic non-robust single-agent strategies. Experiments in this thesis included adapting the task and the action sets to enable agents to converge to robust single-agent soccer strategies. However, the results indicated that the learning of multi-agent cooperative strategies in this environment may be intractable due to an excessively large state-space and the existence of a single-agent policy local-minimum. Furthermore, I argue, via inference from the experiments conducted in this thesis, that the learned policies in another Schmidhuber et al. experiment converge to a similar simplistic single-agent strategy.

[Search These Pages] [DAI Home Page] [Comment]