New AI Method Improves Searchless Chess Play and Puzzle Solving

Current high-performing searchless chess networks, like Chessformer, typically learn by imitating the move choices of stronger, search-based AI systems such as AlphaZero. This paper argues that merely imitating a search is not optimal for playing without one. Instead, the researchers propose fine-tuning these networks using self-play reinforcement learning (RL). A key innovation is replacing the usual entropy bonus for exploration with a "prior-directed exploration" strategy. This new method guides the AI's exploration towards moves that its own Monte Carlo Tree Search (MCTS) prior deems promising. Coupled with an entropy-adaptive sampling temperature that sharpens as a position becomes clearer, this approach led to notable improvements. The system increased puzzle accuracy from 93.9% to 94.9% and mate-in-four accuracy from 77% to 81%, while maintaining or slightly exceeding its base playing strength. The study also highlights that improving puzzle-solving ability does not directly translate to stronger overall play, indicating a dissociation between tactical accuracy and strategic strength.