Improving Monte-Carlo tree search for dots-and-boxes with a novel board representation and artificial neural networks

Dots-and-Boxes is a well-known paper-and-pencil, game for two players. It reaches a high level of complexity, posing an interesting challenge for AI development. Previous, board representation techniques for Dots-and-Boxes rely on data, structures like arrays or linked lists to facilitate operations on the, board. These representation techniques usually lack for the ability, to incrementally update information required for efficient move, generation during search. To address this problem a novel board, representation for Dots-and-Boxes is proposed in this paper. It, utilizes game-specific knowledge to classify distinct conditions on, the board and its implementation is based on disjoint-sets. Besides, the novel board representation this paper treats optimizations for, Monte-Carlo Tree Search (MCTS) focusing on artificial neural, networks. Finally we implemented our proposed approach in a new program called QDab and conducted experiments showing, that the new board representation improves the efficiency of basic, operations on the board by more than 6 times. Further tests, against other implementations show the superior playing strength, of our approach.