2026-10-13T05:27:11Z https://www.scidb.cn/oai

10.57760/sciencedb.46418 2026-08-21T17:15:15Z

2026-08-21 Reinforcement Learning Agent Playing Go-Moku with Neural Network doi:10.57760/sciencedb.46418 en - Build a rule-based AI player that prioritizes forming five-in-a-row, creating open four, and making live three, while detecting whether the opponent has potential threats such as five-in-a-row, open four, or live three to block. This is used to generate imitation learning data and serve as an opponent for reinforcement learning.- Design a DQN agent that clones the behavior of the rule-based player using game data, then plays against the rule-based agent via reinforcement learning. The agent receives rewards for winning, with extra rewards given for generating live three, live four, and double live three patterns. These experiences are stored in a replay buffer and continuously replayed during training.- Save the trained DQN model to disk and employ Monte Carlo tree search combined with prior knowledge and evaluation networks to look ahead 5–7 moves, significantly improving move quality.- Develop a FastAPI web-based gaming service supporting human vs. AI matches, allowing users to switch between different AI types and adjust search depth. Reinforcement Learning; Agent ; Neural Network wang rui PUBLIC https://creativecommons.org/licenses/by-nd/4.0 dataset Science Data Bank