Authors: Yejin Lin; Jiayi Wen
Addresses: Lab of Intelligent Marine Vehicles of DMU, Dalian Maritime University, Dalian, 116033, China ' Lab of Intelligent Marine Vehicles of DMU, Dalian Maritime University, Dalian, 116033, China
Abstract: The natural deep Q-network (DQN) usually requires a long training time because the data usage efficiency is relatively low due to uniform sampling. Importance sampling (IS) can promote important experiences and is more efficient in the neural network training process. In this paper, an efficient learning mechanism using the IS technique is incorporated into duelling DQN algorithm, and is further applied to path planning task for an agent. Different from the traditional DQN algorithm, proposed algorithm improves the sampling efficiency. In this experiment, four target points on the map are deployed to evaluate the loss and the accumulated reward. Simulations and comparisons in various simulation situations demonstrate the effectiveness and superiority of the proposed path planning scheme for an intelligent agent.
Keywords: path planning; DQNs; deep Q-networks; reinforcement learning; importance sampling.
International Journal of Vehicle Design, 2023 Vol.91 No.1/2/3, pp.232 - 247
Received: 11 Feb 2022
Received in revised form: 24 Aug 2022
Accepted: 25 Aug 2022
Published online: 22 May 2023 *