Opponent modeling based on action table for MCTS-based fighting game AI
Man-Je Kim, Kyung-Joong Kim · 2017
Recently, there has been much interest in real-time game AI but it has suffered from short response time with uncountable game complexity. If a forward model is available, the Monte-Carlo Tree Search (MCTS) can also be used for the real-time video games. For example, MCTS has dominated the winning entries in the Fighting game AI competitions. However, because of the response time limitation, their MCTS simulates only five randomly selected actions on the opponent side. Although it works, it's likely to produce outcomes ignoring opponent's playing patterns. In this paper, we propose to incorporate the opponent action prediction based on action table into the MCTS. The AI updates the table during game matches against the opponent. Experimental results show that the approach can help to improve performance against the top three AIs from 2016 IEEE CIG Fighting game AI competitions.