IMPLEMENTATION OF REINFORCEMENT LEARNING AGENT IN A TURN-BASED STRATEGY GAME

• Main Author: Jonathan, Gabriel
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76874
• Institution: Institut Teknologi Bandung

Lack of adaptability has long been recognized as a significant challenge in turn­ based strategy game agents. This is due to the prevalent use of heuristic-based approaches in their decision-making. In this research, an agent trained by Reinforcement Learning is used as a substitute for heuristics in an attempt to overcome this limitation. The agent is trained using the Proximal Policy Optimization algorithm, as implemented in Unity's ML-Agents Toolkit to play testbed game with similar gameplay mechanics to that of a commercial turn­ based strategy game. The pe,jormance of the agent is benchmarked against two systemic approaches, Monte Carlo Tree Search and Online Evolutionary Planning, which have been validated to be appropriate substitutes for heuristics in prior work. Experiments show that the Reinforcement Learning agent encounter challenges pertaining to adaptabili ty, particularly in generalizing opponent behavior that significantly deviates from its training data. However, the performance of the Reinforcement Learning agent is competitive to that of the Monte Carlo Tree Search and Online Evolutionary Planning agents, with an average win rate of 53.7%, showing that Reinforcement Learning can be a promising approach with further improvement and optimization in future work.