Application Of Artificial Potential Field Algorithm Optimized By Deep Reinforcement Learning In Indoor Escape Path Planning

Indoor fires are complex and changeable.When dealing with a completely unknown and complex fire environment,the traditional artificial potential field method is difficult to successfully complete the task of path planning due to its own defects.In recent years,deep learning and reinforcement learning have been constantly moving forward.Using deep reinforcement learning to implement agent path planning tasks has always been a frontier hot research.When the traditional artificial potential field method is applied to a complex and unknown environment,it will fail to find the path due to unreachable targets or local extreme points.And deep reinforcement learning is to allow the agent to learn related strategies to avoid obstacles and find the target point in the process of constantly "making mistakes",and finally obtain an optimal path.Therefore,this subject has carried out the research problem of the traditional path planning algorithm artificial potential field method based on deep reinforcement learning optimization.The main research of the thesis is as follows:(1)This article uses deep reinforcement learning to optimize the artificial potential field method,and according to the difference of deep reinforcement learning classification,respectively proposes a strategy-based artificial potential field optimization method(DDPG-APF)and a value-based artificial potential field optimization method(DQN-APF).Design a special reward function.When the agent is in trouble in the potential field,guide the agent out of the trouble by continuously giving the agent a negative reward value;the DRL-APF algorithm in different scale scenarios requires repeated training and consumes a lot of computing power Under the circumstances,this paper proposes to introduce the mechanism of transfer learning in the DRL-APF algorithm,split the large-scale scene into several small-scale scenes,and apply the learned strategies to the large-scale scenes through the learning of the smallscale scenes to improve The training efficiency of the DRL-APF algorithm.(2)The experiment selected Harbin Geographic Information Industrial Park in Heilongjiang Province as the research area for environmental simulation experiments.The results show that both types of optimization algorithms can achieve path planning in complex environments;under two random obstacle environments,strategy-based optimization The method(DDPG-APF)requires 13.2 seconds less time to plan the complete path than the value-based optimization method(DQN-APF),and the efficiency is 24.4% higher;to prove the generalization of the algorithm,random obstacles are increased to 4 At 1 and 6,the time required for DDPG-APF to plan the complete path is 7.2 seconds and 10.1 seconds less than that of DQN-APF,respectively,and the efficiency is 11.7% and 15.9% higher,which shows that the two optimization algorithms have good generalization capabilities..(3)The DRL-APF algorithm path planning training efficiency experiment before and after the introduction of transfer learning in different scale scenarios has been carried out.The experiment shows that before the introduction of transfer learning,the agent is trained in two scenarios of 60m*90m and 70m*70m,40,000 The round reward value has not yet converged,and after the introduction of transfer learning,the reward value obtained by the agent in the two scenarios converged before 20000 rounds.It proves the feasibility of introducing transfer learning into the DRL-APF algorithm in this paper.