WMGIRL: Research On Multi-Agent Inverse Reinforcement Learning Algorithms And Their Applications In Transportation

With the continuous improvement of my country's comprehensive national strength,many cities have successively hosted large-scale events,and urban roads cannot meet the rapidly increasing traffic demand in a short period of time.Therefore,the traffic police will use traffic control schemes to induce traffic to relieve road pressure.In view of the above situation,this paper proposes a multi-agent inverse reinforcement learning(WMGIRL)algorithm based on intra-group evolution to predict the flow changes on the urban road network under the implementation of traffic control schemes and visualize Display to provide data support for traffic police decisionmaking.The WMGIRL algorithm is improved based on the maximum entropy inverse reinforcement learning algorithm and the DDPG algorithm.The main work of this article is as follows:First,the road network approach simplifies network-based: the environment used to build the model of inverse reinforcement learning,the method can crawl and simplify the real city road network,urban and road network data and bayonet match.First of road network modeling,extracting the connection between the waypoints as the basis for a simplified road network,road network and de-noising and streamline operations,not only reduces the workload for the maximum entropy algorithm order to solve the environment model problem of excessive and does not undermine the relationship between the inner link road network intersection.Second,the learning algorithm Reverse strengthen based on maximum entropy weight and more weight: The algorithm uses the inverse reinforcement learning expert strategy set as the research object characteristics of the environment reward function of digging,digging out a unique driving track characteristics of urban residents.Maximum Entropy traditional inverse reinforcement learning algorithm inadequate discrimination on the critical state influence policy experts in special environments and the accuracy of the algorithm,the algorithm is weighted policy experts,key state and environment marked,in order to improve convergence rate.Third,the multi-agent discriminative multi-agent reinforcement learning algorithm based on intra-group evolution: used to replace the forward-backward algorithm in maximum entropy inverse reinforcement learning.This algorithm is a multi-agent reinforcement learning algorithm based on DDPG.communicate autonomous decision-making training,and the introduction of mechanisms such as genetic algorithm convergence within the group to improve efficiency.Fourth,build a traffic simulation platform for variable traffic control schemes:This article is to build a traffic simulation platform based on real road network on the basis of the proposed algorithm on.The platform can be deployed in two different types of traffic control program.This article uses real traffic data to verify the accuracy of algorithms and platforms.Experimental results show that the region has a high accuracy traffic distribution algorithms build a platform in this paper predicted in the traffic control program.