ETQ-learning: an improved Q-learning algorithm for path planning

• Published in: Intelligent service robotics Vol. 17; no. 4; pp. 915 - 929
• Main Authors: Wang, Huanwei, Jing, Jing, Wang, Qianlv, He, Hongqi, Qi, Xuyan, Lou, Rui
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.07.2024; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Collision avoidance; Control; Decision making; Design; Dynamical Systems; Engineering; Greedy algorithms; Machine learning; Mechatronics; Obstacle avoidance; Optimization; Original Research Paper; Path planning; Planning; Redesign; Robot dynamics; Robotics; Robotics and Automation; Robots; Strategy; Unmanned aerial vehicles; User Interfaces and Human Computer Interaction; Vibration; Q-learning; Path planning; Greedy strategy; Reward mechanism; Reinforcement learning
• ISSN: 1861-2776; 1861-2784
• DOI: 10.1007/s11370-024-00544-3

Path planning algorithm has always been the core of intelligent robot research; a good path planning algorithm can significantly enhance the efficiency of robots in executing tasks. As the application scenarios for intelligent robots continue to diversify, their adaptability to the environment has become a key focus in current path planning algorithm research. As one of the classic reinforcement learning algorithms, Q-learning (QL) algorithm has its inherent advantages in adapting to the environment, but it also faces various challenges and shortcomings. These issues are primarily centered around suboptimal path planning, slow convergence speed, weak generalization capability and poor obstacle avoidance performance. In order to solve these issues in the QL algorithm, we have carried out the following work. (1) We redesign the reward mechanism of QL algorithm. The traditional Q-learning algorithm’s reward mechanism is simple to implement but lacks directionality. We propose a combined reward mechanism of "static assignment + dynamic adjustment." This mechanism can address the issue of random path selection and ultimately lead to optimal path planning. (2) We redesign the greedy strategy of QL algorithm. In the traditional Q-learning algorithm, the greedy factor in the strategy is either randomly generated or set manually, which limits its applicability to some extent. It is difficult to effectively applied to different physical environments and scenarios, which is the fundamental reason for the poor generalization capability of the algorithm. We propose a dynamic adjustment of the greedy factor, known as the ε - a c c - i n c r e a s i n g greedy strategy, which significantly improves the efficiency of Q-learning algorithm and enhances its generalization capability so that the algorithm has a wider range of application scenarios. (3) We introduce a concept to enhance the algorithm’s obstacle avoidance performance. We design the expansion distance, which pre-sets a "collision buffer" between the obstacle and agent to enhance the algorithm’s obstacle avoidance performance.