Research On Electric Vehicle Routing Problem Based On Reinforcement Learning

With the rapid development of our national economy,the social logistics demand is higher and higher,but the ratio of total social logistics costs to GDP remains high.And transportation costs are more than half of social logistics costs,so the research on making vehicle path planning has important practical value and strategic significance.Due to the requirements of low carbon economy and the popularity of green logistics concept,however,substituting conventional fuel vehicles with low-energy and low-pollution electric vehicles is the trend of times,which has important meanings in both energy industry and environmental protection.According to the features and the restrictions of electric vehicle delivery task,this dissertation abstracts out the electric vehicle routing problem(EVRP)and constructs a mathematical model to minimize the total delivery cost with constraints of time,load and battery capacity based on social charging network.Then this dissertation designs an algorithm RL-EVRP to solve EVRP based on Actor-Critic.The algorithm,consisting of an Actor network and a Critic network,generates instances sampled from a given distribution and trains a model by the Adam algorithm.The trained model is able to solve other instances from similar distribution at speed due to no need to re-train.In order to ensure solutions feasible,this dissertation designs the corresponding masking scheme at decoding phase in Actor network.Lastly,simulation results show that the RL-EVRP can get shorter total delivery cost and fewer vehicles and that the reinforcement learning can be applied to complicated combinatorial optimization problem successfully.