Limitations in Planning Ability in AlphaZero

Abstract

AlphaZero, a deep reinforcement learning algorithm, has achieved superhuman performance in complex games like Chess and Go. However, its strategic planning ability beyond winning games remains unclear. We investigated this using 4-in-a-row, a game used to study human planning. We analyzed AlphaZero's feature learning and puzzle-solving abilities. Despite strong gameplay, AlphaZero exhibited a 93% failure rate in puzzles. Our feature analysis showed that its self-learned strategies during training lacked certain critical human-like features. We added human-inspired cognitive value function into its policy and value outputs, leading to a 15% improvement in puzzle-solving accuracy. Our findings highlight the potential for human insights to enhance AI's strategic planning beyond self-play.