A novel real-time design for fighting game AI

Real-time fighting games are challenging for computer agents in that actions must be decided within a relatively short cycle of time, usually in milliseconds or less. That is only achievable by either very powerful machines or state-of-the-art algorithms. The former is usually a costly option while the latter remains an ongoing research topic despite countless research. This paper describes our algorithmic approach towards real-time fighting games via the fighting game AI challenge. The focus of our research is the LUD division, the most challenging category of the competition where action data is hidden to prevent methods that are dependent on prior training. In this paper, we propose several generic heuristics that can be used in combination with Monte-Carlo tree search. Our experimental results show that such an approach would provide an excellent AI outperforming pure Monte-Carlo tree search and classic algorithms such as evolutionary algorithms or deep reinforcement learning. Nonetheless, we believe that our proposed heuristics should be able to generalize to other domains beyond the scope of the fighting game AI challenge.