Learning-Based Decision Making in a Competitive Card Game
Author
Loós, Tamás
Term
4. term
Education
Publication year
2026
Submitted on
2026-02-24
Pages
46
Abstract
This thesis investigates behavioral cloning for learning to play Scripts of Tribute, a competitive deck-building card game used as a testbed at the IEEE Conference on Games AI Competition. A neural network was trained to imitate an expert MCTS bot using 6,400 games (673,619 decisions), achieving approximately 59% top-1 accuracy, approximately 15% above the strongest trivial strategy. This accuracy plateau persisted across five model configurations (84K-2.3M parameters) and two training methods (behavioral cloning and DAgger), suggesting the bottleneck is not model capacity but the teacher's reliance on search information unavailable to the student. A pointer-based architecture addresses a mismatch between positional action encoding and input processing that ignores card order, scoring actions by card content rather than position. The deployed bot (457K parameters) makes decisions in approximately 5ms and achieves 36.9% overall win rate against eleven competition bots, beating heuristic opponents but winning only 12-15% against the strongest MCTS-based bots. This gap is consistent with the compounding-error problem of behavioral cloning.
Keywords
Documents