실시간 Q-learning을 이용한 모바일 로봇의 최적 경로 탐색

Abstract

Recently, as the field of autonomous driving has been actively researched, the importance of route search is increasing. In particular, reinforcement learning is known to be useful for sequential decision-making problems, and research is focused on autonomous driving using reinforcement learning. Recently, it is producing successful results in fields that have not been solved until now in a form combined with a neural network. However, this requires very complex algorithms and high cost. In this paper, path search was implemented using Q-learning, a simple reinforcement learning algorithm. However, Q-learning is not suitable for dynamic environments with infinite states as it requires training the Q-table in advance for each state. To overcome this limitation of Q-learning, real-time Q-learning was used. To use real-time Q-learning, it was necessary to increase the learning rate, and it was satisfactory by adjusting the search strategy and reward. To show that real-time Q-learning is useful, we compared it with DQN and showed significant performance. Finally, we simulated something capable of responding to dynamic obstacles and applied it to a real mobile robot.