Komparasi Performa Algoritma Q-Learning dan Algoritma SARSA dalam Perancangan Agen pada Permainan Catur Jawa

Abstract

Reinforcement learning is a prominent method in the development of artificial intelligence for computer games, but its application in board games, especially traditional Indonesian board games, is still relatively limited. In reinforcement learning, there are two factors that influence algorithm performance, namely learning rate and discount factor, which determine the optimal model to produce the best intelligent agent. This research compares two reinforcement learning algorithms, Q-Learning and SARSA, in Javanese chess, by comparing the winning rate of each agent using various combinations of learning rate and discount factor values, and determining the combination of these two parameters that is suitable for an agent to win a round of a match. The test results show that in the scenario of the Q-Learning algorithm against SARSA or vice versa, the SARSA algorithm outperforms Q-Learning with a win rate of 58.4% compared to 58.1% for agent P1. Meanwhile, for P2 agents, the Q-Learning algorithm outperformed SARSA with a win rate of 28.9% versus 28.8% of the total win percentage. The optimal parameters for Q-Learning are a learning rate of 0.06 and a discount factor of 0.8 for agent P1, and 0.08 and 0.1 for agent P2. For the SARSA algorithm, the optimal parameters are 0.03 and 0.5 for agent P1, and 0.06 and 0.7 for agent P2. These findings provide valuable insights into selecting intelligent agents in board games other than Javanese chess, thereby contributing to the advancement of Artificial Intelligence in the context of these games.