Reinforcement Learning for Turn-Based Strategy Game

Predictability and lack of adaptability have long been recognized as significant challenges in strategy turn-based games. This is due to the prevalent use of heuristic-based approaches in their agent's decision-making. In this paper, we attempt to solve this issue by replacing heuristics with an intelligent agent trained by Reinforcement Learning (RL). We compare the performance of our trained agent against two systemic approaches, Monte Carlo Tree Search (MCTS) and Online Evolutionary Planning (OEP). Experiments show that while RL is vulnerable to play styles that differ sharply from its training data, it is competitive against MCTS but falls short to OEP. This suggests that RL can be a promising approach after further improvement and optimization in future work.