Amini, Omid2025-04-292025-04-292025-04-29https://hdl.handle.net/10222/85065This thesis investigates how advanced machine-learning techniques can rank and compare player behaviours in RoboCup Soccer Simulation 2D (SS2D), surpassing traditional metrics such as goals and assists. We fused official RoboCup logs with large-scale simulated matches to build robust data sets. Then, we used Variational Auto-Encoders and Transformers to compress high-dimensional state–action sequences into low-dimensional embeddings that preserve essential spatio-temporal patterns. We applied Maximum-Entropy Inverse Reinforcement Learning to uncover each player’s reward function, producing context-aware scores that capture both on-ball and off-ball actions. A complementary win-probability learner (“WinGain”) linked individual actions to match success. We tested framework robustness across various match conditions (communication bans, player swaps and formation changes). Our framework uncovers effective behaviours overlooked by raw statistics, enables objective cross-player rankings, and provides a data-centric foundation for optimizing strategies in SS2D and other complex multi-agent sports domains, which can aid coaches, scouts and AI researchers in evidence-based decision-making.enPlayer behaviour analysisRoboCup Soccer Simulation 2DMachine learning in sports analyticsMulti-agent systemsReward-based player evaluationVariational Auto-Encoder (VAE)Reward-based player evaluationTransformer embeddingsInverse Reinforcement Learning (IRL)