An Improved Path Planning Algorithm Based on Densely Connected Convolutional Network and Dueling Network Architecture
HUANG Ying
YU Yuqin
Abstract：When the existing deep Q-networks are used in path planning domain,they surfer with overestimations of ac?tion-state values and can't meet the need for real-time path planning. So,an improved path planning algorithm based on densely connected convolutional network and dueling network architecture is proposed. Firstly,a network that fused simplified densely con?nected convolutional network with dueling network architecture is proposed,which is a lighter deep network for deep reinforcement learning. Then reinforcement learning methods are used to solve the path planning problem and train the proposed network by double deep Q-network(DDQN)algorithm to approximate the optimal action-state value function. Finally,experiments in the customized gridmap environment are done. Experiments demonstrate that our proposed algorithm can not only obtain less parameters,computa?tion time and lower training expense,and meet the need for real-time path planning,but also can lead to more state-of-the-art per?formance on route planning domain which can increase the path planning success rate by about 5% on average with great generaliza?tion ability of rapidly changing environments.
Keywords：deep reinforcement learningpath planningdensely connected convolutional networkdueling network archi?tecturedouble deep Q-network
Publication Date：2019-01-01
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：8( 812-819 )