Adaptation by prediction: Reading the play in robot soccer

David Michael Ball (2006). Adaptation by prediction: Reading the play in robot soccer PhD Thesis, School of Information Tech & Elec Engineering, The University of Queensland.

       
Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
n33692430_phd_abstract.pdf n33692430_phd_abstract.pdf application/pdf 6.92MB 14
n33692430_phd_content.pdf n33692430_phd_content.pdf application/pdf 6.91MB 1
n33692430_phd_front.pdf n33692430_phd_front.pdf application/pdf 6.92MB 59
n33692430_phd_totalthesis.pdf n33692430_phd_totalthesis.pdf application/pdf 6.90MB 7
Author David Michael Ball
Thesis Title Adaptation by prediction: Reading the play in robot soccer
School, Centre or Institute School of Information Tech & Elec Engineering
Institution The University of Queensland
Publication date 2006-11
Thesis type PhD Thesis
Supervisor Gallagher, Marcus R.
Wyeth, Gordon F.
Subjects 290000 Engineering and Technology
Formatted abstract This thesis describes adaptation by modelling and prediction for autonomous agents in spatial, dynamic, and multi-agent environments. The study of agents in these environments is relevant to applications ranging from the autonomous operation of robots to the design of non-player characters in computer games. For these applications, it is typical to assess the state of the environment to select appropriate actions for the agent(s). This thesis explores whether predicting and adapting to the likely future state of the environment – “reading the play” – can improve the agents’ performance.
The studies in this thesis are based in the RoboCup robot soccer environment which is representative of the challenging environments addressed by this research. Robot soccer agents operate in a three degree of freedom space with ten agents moving at speeds over one metre per second and a ball that sometimes travels at up to five metres per second. The key results from the studies are derived from statistical analysis of many games using a high fidelity simulation of a world leading real robot system.
The first study, adaptation by prediction of agent behaviour, explores classifying behaviour in a spatial and dynamic environment. The basis of the approach is to predict the opponent planner’s assignment of roles, and exploit them based on an understanding of these roles. The research compared a supervised learning Naïve Bayesian Classifier with an expert defined Fuzzy Classifier approach and the results show that both methods can classify behaviour.
The second study, adaptation by prediction of agent motion, has two parts. The first part explores the development of a module to predict an agent’s likely future occupancy using Markov Chains. Occupancy Grids can represent an agent’s probabilistic grid position; this thesis describes a method of extending occupancy grids to represent and capture an agent’s motion, termed “Motion and Occupancy Grids.” The thesis describes an efficient algorithm to dynamically build the Markov State Transition Matrix that links an agent’s current state to its likely future state. The results from a range of experiments demonstrate excellent performance in predicting the agent’s future occupancy.
The second part studies the value of these predictions of agent occupancy by integrating them into a planning system. An agent can potentially select better roles by knowing the other agent’s future state. The experiments compare the performance of the planning system without predictions to the planning system using the predictions as a resource. The results of the study demonstrate a surprising decrease in performance when using the predictions of agent occupancy. A consistent pattern of poor performance was found across a range of planner parameter settings which is illustrated through a case study of the best settings.
The last study, adaptation by modelling of behaviour patterns, explores the connection between the adaptation approaches and the planner methodology using statistical modelling of behaviour patterns. The basis of the approach is to model potentially useful characteristics of agent group behaviour and to make plans that exploit the understanding of these patterns. The results show that the performance increases when the model and adaptation approach bias the solution which closely matches the goal of the planner.
The overall conclusion of the research is that to improve the performance of an autonomous agent system the agent designer must carefully match the adaptation method with the heuristics of the planner. Intuitively it seems that performance would increase by using more accurate and expressive representations of the likely effect of the opposition. However, indecision and oscillations are the result of naïve planning that does not consider the relationship between the action space and the predictions. Increasing performance by adaptation is possible in dynamic and spatial environments, demonstrated by statistical modelling of the opponent team’s behaviour patterns, when appropriately matched to the planner’s underlying principles.


 
Citation counts: Google Scholar Search Google Scholar
Access Statistics: 471 Abstract Views, 81 File Downloads  -  Detailed Statistics
Created: Wed, 20 Aug 2008, 11:50:21 EST by Noela Stallard on behalf of Library - Information Access Service