AI and real-time Strategy Games

Abstract: As artificial intelligence continues to progress, applications hope to have AI perform tasks as good, if not better, than a human. A good way to test theoretical applications is through games, simple to complex. Artificial intelligence models have been used in real-time strategy games over the past few years, but their implementations remain rudimentary, and there is much work to be done. The research question is how can Monte Carlo, a prominent algorithm in today’s world, can be improved both generally and in the context of real-time strategy games specifically. Implementing an experimental design is the main research methodology. This technique was selected because it offers the clearest framework for contrasting to unexplored ideas. The research question is based around improving the Monte Carlo method, particularly in MicroRTS, a popular test environment for AI algorithms. As the goal of the research is to enhance the Monte Carlo Tree Search (MCTS) algorithm for strategy games, an evaluation of conventional MCTS implementations (MCTS Greedy and MCTS UCB), and a unique suggested MCTS (MCTS UCB+) will be done in comparison. Benchmarking each algorithm’s performance in an RTS environment and comparing the results based on several criteria is how the comparative analysis is accomplished. The study finds that through changes to the sampling and selection methods and the understanding of the game state, the new algorithm MCTS UCB+ is able to outperform its predecessors in the partially observable game mode of MicroRTS.