Research On Dynamic Path Planning By Evolutionary And Reinforcement Learning Algorithms

With the rapid development of robot technology,a large number of robots have been applied to military,commercial and people's livelihood fields,the intelligent control of robots has become a research hot spot.In the dynamic path planning environment requires the intelligence to have the ability to make real-time decisions on the changes of the environment,The deep learning-based reinforcement learning algorithm performs well in complex environmental tasks that are difficult to complete by traditional methods.Nowadays,deep reinforcement learning has been widely used to solve intelligent control problems,However,deep reinforcement learning method usually has three core problems: time credit allocation with sparse rewards,lack of effective exploration,and extremely sensitive convergence to super parameters.For these problems,this paper proposes an improved swarm intelligence deep reinforcement learning algorithm.The setting of the evaluation function of swarm intelligence algorithm can directly judge the quality of the final result,which has a good performance in the problem of time credit allocation.At the same time,this method can also make the experience collected by the algorithm tend to be of high long-term return,which makes the exploration experience more convenient for training.Swarm intelligence algorithm only needs to consider the problem of search time and can reduce the dependence of algorithm convergence on super parameters.Experimental results show that this method has certain theoretical and practical value.Main research contents:(1)An improved genetic neural network reinforcement learning algorithm is proposed in this paper.The adaptive function conforming to the characteristics of reinforcement learning was proposed in the algorithm,and the method of segmental weight selection crossover was proposed to breed offspring and the method of using gradient information to accelerate the search of solution space of the genetic algorithm.The performance of the algorithm was verified in the modified sparse return deep reinforcement learning experimental environment.(2)According to the idea of swarm optimization,another reinforcement learning algorithm of particle swarm optimization neural network is proposed.In order to deal with the pseudo-optimal characteristics of networks in deep reinforcement learning,a global historical table with the first few particles(individuals who rate the first few of all historical individuals)is proposed to make the algorithm more stable.Finally,the performance of the algorithm is verified in the modified sparse reward deep reinforcement learning experimental environment.(3)The algorithm is verified experimentally in the dynamic path planning problem.For the dynamic path planning environment of a continuous action space,the non-sparse and sparse cases are verified respectively.The experimental results show that the deep reinforcement learning method based on swarm intelligence optimization can effectively solve the problem of spatial path planning of continuous motion with different forms of reward.