Agent Performance Comparison of the Q-Learning Algorithm and SARSA Algorithm in Javanese Chess Game

Reinforcement learning is a prominent method in Artificial Intelligence development for computer games, but its application in computer-based board games, particularly traditional ones, is still limited. In reinforcement learning, there are two key factors influencing algorithm performance, namely learning rate and discount factor, which determine the optimal model to produce the best intelligent agent. This study aims to compare two reinforcement learning algorithms, Q-Learning and SARSA, in a Javanese chess game. Specifically, we investigate how different combinations of learning rates and discount factors influence the winning rates of agents. We conducted tests across 100 rounds for each parameter combination variation. The results showed that the SARSA agent outperformed the Q -Learning agent with a winning rate of $\mathbf{5 9. 6 \%}$ against $52.8 \%$ for P1 agents. Meanwhile, for P2 agents, the Q-learning agent outperformed the SARSA agent with winning rates of $\mathbf{2 8. 4 \%}$ against $\mathbf{2 8 \%}$, respectively. The optimal parameters identified for the $\mathbf{Q}$-Learning agent included a learning rate of 0.06 and a discount factor of 0.8 for P1 agents and 0.02 and $\mathbf{0. 4}$ for $\mathbf{P 2}$ agents. Meanwhile, for the SARSA agent, optimal parameters were 0.07 and 0.6 for P1 agents and 0.04 and 0.5 for $\mathbf{P} 2$ agents. These findings provide valuable insights for selecting intelligent agents in board games beyond Javanese chess, contributing to the advancement of Artificial Intelligence in such gaming contexts.