Date of Award
2025
Document Type
Thesis
Publisher
Edith Cowan University
Degree Name
Doctor of Philosophy
School
School of Science
First Supervisor
Martin Masek
Second Supervisor
Chiou Peng Lam
Abstract
Developing artificial intelligence (AI) systems to automate the behaviour discovery process for agents in multi-agent systems (MAS) requires significant manual effort, domain expertise, and iterative testing. This results in high costs and substantial resource demands. Real-time strategy (RTS) games, a type of MAS characterised by competing teams of agents in dynamic environments, serve as a well-studied medium for exploring multi-agent behaviour discovery. Traditional methods for AI behaviour discovery have been applied to RTS games but can introduce human biases, limiting the exploration of novel strategies and reducing the richness of player experiences. To address these challenges, this thesis presents an approach that integrates cooperative co-evolutionary algorithms (CCEA) with a dynamic decomposition scheme to automate the discovery of team behaviours in RTS games.
By dynamically partitioning the problem space based on a combination of measures from gameplay and the evolutionary process the proposed approach minimises manual human intervention while optimising team behaviours across diverse gameplay scenarios. Central to the investigation is the study of novel decomposition operators that dynamically modify how the problem is partitioned during the execution of the algorithm, facilitating the emergence of co-operative behaviours among agents without requiring expert knowledge. This adaptability reduces human biases and enhances the performance of the behaviour discovery process.
Experimental results demonstrated that the approach consistently discovered winning behaviours, outperforming traditional static decomposition methods across a range of studies with a varying number of agents and scales of environmental complexity. The approach’s scalability and effectiveness in complex, large-scale environments highlight its potential for broader applications in MAS and related domains.
To evaluate the approach outside of the RTS domain, it was applied to an image processing problem, generating visual counterfactual explanations for medical images. This extension showcases the adaptability of the dynamic decomposition approach beyond MAS, with potential application for operations research and other fields that involve spatial problem-solving.
This thesis contributes a novel approach and perspective for simultaneous problem decomposition and optimisation. It enables bias-free problem-solving approaches, extends knowledge in agent-based simulation scalability, and presents a generalisable dynamic decomposition method applicable to diverse problem domains. Collectively, these contributions enhance the ability to solve complex problems in MAS and related fields.
DOI
10.25958/wmt8-7215
Access Note
Access to this thesis is embargoed until 22nd March 2026
Recommended Citation
Kelly, L. B. (2025). A cooperative co-evolutionary approach to the automated discovery of team behaviour in real-time strategy games. Edith Cowan University. https://doi.org/10.25958/wmt8-7215