Action-selection in RoboCup keepaway soccer : experimenting with player confidence : a thesis presented in partial fulfilment of the requirements for the degree of Masters of Science in Computer Science at Massey University
Through the investigation of collaborative multi-agent domains, in particular those of robot soccer and robot rescue, and the examination of many popular action-selection methodologies, this study identifies some of the issues surrounding entropy, action-selection and performance analysis. In order to address these issues, a meaningful method of on-field player evaluation, the confidence model, was first proposed then implemented as an action-selection policy. This model represented player skill through the use of percentages signifying relative strength and weakness and was implemented using a combination of ideas taken from Bayesian Theory. Neural Networks. Reinforcement Learning, Q-Learning and Potential Fields. Through the course of this study, the proposed confidence model action-selection methodology was thoroughly tested using the Keepaway Soccer Framework developed by Stone, Kuhlmann, Taylor and Liu and compared with the performance of its peers. Empirical test results were also presented, demonstrating both the viability and flexibility of this approach as a sound, homogeneous solution, for a team wishing to implement a quickly trainable performance analysis solution.