Mimicking the Human Approach in the Game of Hive

While Deep Blue and AlphaGo make it seem like board games have been solved, there are still plenty of games for which no good game playing program exists. Hive is such a game. It is, combinatorically, of similar complexity as chess or Go, yet the rules of the game are such that current methods can barely beat a randomly playing agent. A major bottleneck for progress is the high branching factor of the game. We apply state of the art methods for which we develop new heuristics that are based on human domain-knowledge, attempting to improve upon the current dire state of Hive agents. Our methods have improved playing strength compared to the state of the art, although our AI still fails against actual Humans. We also find that, while in most board games, brute force or deep learning approaches work best, in Hive, an approach based on mimicking human knowledge outperforms these other approaches, including Monte Carlo Tree Search or Deep Reinforcement Learning. Future work will show if this anomalous situation is inherent to the game.