Entropy-Guided Exploration in AlphaZero: Enhancing MCTS with Information Gain for Strategic Decision Making
Shuqin Li, Xueli Jia, Yadong Zhou, Dong Bin Xu, Haiping Zhang, Miao Hu · 2024
AlphaZero has revolutionized AI game-playing, yet it may overlook actions crucial for long-term strategy. This paper introduces Entropy-Guided MCTS (EG-MCTS), enhancing AlphaZero by incorporating information theory into action selection. We propose an information gain metric based on game state entropy, integrating it into MCTS to balance immediate rewards with informational value. Our contributions include developing this metric, modifying neural network training, and systematically balancing information gain with existing criteria. Comprehensive evaluations across various games demonstrate EG-MCTS's improved performance in complex strategic scenarios, showing enhanced efficiency in game tree navigation and adaptation to unfamiliar situations. These findings suggest broader applications in sequential decision-making under uncertainty, advancing the development of more adaptable AI systems.