Speakers
Description
This paper presents a comparative study of different neural network architectures trained using behavioral learning in the strategic board game Blood Bowl. The game’s complexity, driven by its large branching factor and inherent randomness, presents a significant challenge for artificial intelligence (AI). Traditional AI approaches, such as scripted and search-based methods, have struggled to achieve human-level performance. This study examines how various deep learning models process decision data, adapt strategies, and handle uncertainty in gameplay. Our methodology involves training models with a custom preprocessing pipeline that extracts valid game states and actions from replays of scripted solutions and tournament self-play. Performance is evaluated by competing against existing AI solutions, using metrics such as win rates, move efficiency, and strategic accuracy. The results highlight the comparative strengths and weaknesses of each architecture, providing insights into their effectiveness for reinforcement learning-based agents in complex decision-making environments like Blood Bowl.
