Fuzzy Rule based Generative Adversarial Imitation Learning

초록

Reinforcement learning is a crucial technology for implementing intelligent service scenarios by optimizing reward functions to learn similar action. However, reward functions designed by humans often lead to significant errors when imitating action. To address this issue, methods for reward estimation and behavior generation through inverse reinforcement learning have been explored. Nevertheless, Generative Adversarial Imitation Learning(GAIL) can suffer from decreased learning speed and performance in uncertain environments with limited information. We propose designing the Generator of GAIL based on fuzzy rules, aiming for faster and more efficient action generation. This approach focuses on improving the algorithm's learning speed and preventing performance degradation in uncertain environments. Simulation results demonstrate that the proposed method outperforms existing algorithms in terms of stability and performance. © 2025 IEEE.

키워드