What question did this study set out to answer?

This research aims to develop strategies for perfect information games using a novel learning method to optimize action choice.

May 2, 2026

Generating Strategies for Games with Complete Information Using Action Similarity

Key Points

This research aims to develop strategies for perfect information games using a novel learning method to optimize action choice.
Implemented a probabilistic-combinatorial formal learning method alongside traditional Monte Carlo evaluations.
Evaluated generalized rules to generate strategies based on state-action similarities.
Explored the relationship between action similarities and rewards in gameplay.
Demonstrated that the new method produces more effective strategies compared to traditional approaches.
Achieved a notable improvement in generalization ability for strategy generation.
Identified successful patterns in states corresponding to high reward actions.

Abstract

In generating (sub-)optimal strategies for perfect information games, the dominant paradigm is reinforcement learning using neural networks to estimate actions (Q-values). The initially applied approach using a tree search with Monte Carlo evaluations was abandoned due to its lack of generalization ability. By applying a probabilistic-combinatorial formal learning method together with the Monte Carlo method we will show how generalized rules can be generated that form the desired strategy through similarities in states where the same action was applied leading to high reward.

Bookmark

Generating Strategies for Games with Complete Information Using Action Similarity

Key Points

Abstract

Cite This Study

Also Consider

Also Consider