文献类型：期刊文献

英文题名：Enhanced Autonomous Navigation of Robots by Deep Reinforcement Learning Algorithm with Multistep Method

作者：Peng, Xiaohong[1];Chen, Rongfa[1];Zhang, Jun[1];Chen, Bo[2];Tseng, Hsien-Wei[3];Wu, Tung-Lung[3];Meen, Teen-Hang[4]

机构：[1]Guangdong Ocean Univ, Sch Math & Comp Sci, Zhanjiang 524088, Peoples R China;[2]Lingnan Normal Univ, Coll Informat Engn, Zhanjiang 524048, Peoples R China;[3]Longyan Univ, Sch Math & Informat Engn, Longyan 364012, Fujian, Peoples R China;[4]Natl Formosa Univ, Dept Elect Engn, Huwei 632, Yunlin, Taiwan

年份：2021

卷号：33

期号：2

起止页码：825

外文期刊名：SENSORS AND MATERIALS

收录：SCI-EXPANDED(收录号：WOS:000624403600013)、、EI(收录号：20211210108020)、Scopus(收录号：2-s2.0-85102658372)、WOS

语种：英文

外文关键词：autonomous navigation; obstacle avoidance; deep reinforcement learning; navigation planning

外文摘要：In this paper, we propose a new method to improve the autonomous navigation of mobile robots. The new method combines a multistep update method with a double deep Q-network (MS-DDQN) to realize reinforcement learning (RL) to enhance the navigation ability of mobile robots. The proposed MS-DDQN gives two types of rewards for taking actions: terminal and non-terminal rewards. These rewards are subdivided into several different ones including rewards for arrival and collision (terminal rewards) and rewards for distance, orientation, and danger (non-terminal rewards). With the subdivided reward system, the new method trains mobile robots more effectively to increase their autonomous navigation ability. In the experimental process of this study, a laser range finder was used as the sensor for the mobile robot to perceive the distance information of the obstacle. Experiment results validated the new method's enhanced ability, showing higher success rates (97% on average) than those of other methods such as the double deep Q-network (DDQN), prioritized deep Q-network (DQN), and prioritized DDQN. The higher success rates originated from the sophisticated reward system as the total reward of the proposed method was 7-94% higher than those of the other methods in simulations in five different environments. The learning speed was also improved, reducing the learning time, as the new method completed the learning in fewer episodes. The results of the new model suggest that MS-DDQN enables mobile robots to have higher learning efficiency and generalization ability than conventional deep reinforcement (DRL)-based methods and allows them to navigate autonomously even in unknown complex environments.