Research On Path Planning Of Mobile Robot Based On Reinforcement Learning

Traditional path planning algorithms mostly adopt supervised learning strategy,which not only needs to know a lot of environmental information,but also may fall into local optimization.Reinforcement learning algorithm adopts unsupervised learning strategy,so that robots can interact with the environment as well as accumulate experience in the process of searching,and realize independent optimization.Aiming at the problems existing in path planning of mobile robots based on reinforcement learning,this paper proposes AMD＿Q-learning,E＿PE＿D3QN and Self-Attention-DQN algorithms,which can not only improve the optimization efficiency,but also avoid the dimension disaster and balance the relationship between exploration and utilization.Firstly,this paper studies Q-learning algorithm and its application in path planning,analyzes its existing problems,and puts forward AMD＿Q-learning algorithm.AMD＿Qlearning algorithm uses the optimized artificial potential field to initialize the Q table,which enhances the mobile robot’s perception of the environment in the initial stage of optimization.Multi-step strategy is adopted to enrich the robot’s action set,which reduces the robot’s moving steps and optimizes the optimal path.Design a greedy factor with dynamic adjustment ability,so that mobile robots can better balance the relationship between exploration and utilization.Secondly,we analyze the DQN algorithm and its process in path planning.When grid coordinates are used as the input of deep neural network in the optimization environment,aiming at the deficiency of DQN algorithm based on deep neural network,E＿PE＿D3QN algorithm is designed.On the basis of D3 QN,we propose two strategies,on the one hand,the preferential experience playback mechanism is introduced to reorder the samples in the experience pool,so as to improve the sampling probability of important samples,thus improving the optimization efficiency;on the other hand,the greedy factor is adjusted adaptively by establishing the mapping relationship between reward value and greedy factor,and the exploration and utilization of DQN algorithm are improved.When the image is used as the input of the deep neural network in the optimization environment,a convolution neural network with stronger key feature extraction ability and multi-head self-attention mechanism is designed.Based on this network,a Self-Attention-DQN algorithm is designed.After each convolution layer,a batch normalization algorithm is added to unify the data distribution,speed up the training efficiency,and global average pooling is used instead of full connection layer to reduce network parameters and aovid information overload.Finally,the simulation experiments are carried out respectively in grid environment and map environment.The results show that the algorithm proposed in this paper has better optimization path and higher optimization efficiency compared with the algorithm before improvement.