A Neural Network Optimization for Gobang Game Strategy

摘要

This paper proposes a new neural network algorithm for an optimal strategy of two players' board games. The proposed algorithm decides the best next move by thinking a future state where the both situations of two players are even. By using the proposed algorithm, we show a breakthrough for choosing the best next move of Gobang game called "Renju " or "Go-Moku ". Gobang game is played by two players, on a board of 15 x 15 squares with black and white stones. The proposed algorithm chooses a certain move from the hierarchical trees of all possible moves, and searches a path by modifying a bad move so that the simulation of each player keeps even in the future. This feedback strategy is modeled based on the heuristic thinking of various kinds of professional board game player. The simulation result shows that the proposed algorithm can decide an appropriate move with a shorter computational cost than the conventional greedy search algorithms.