Drone-Assisted Vehicle Distribution Route Optimization

The logistics industry has witnessed rapid growth,and a new logistics distribution mode,combining vehicles and drones,has gained significant attention.This innovative approach brings together the benefits of both vehicles and drones,resulting in faster delivery times and reduced distribution costs compared to traditional logistics distribution methods,and the key to using the combination of vehicles and drones distribution mode to achieve cost reduction and efficiency of logistics is to propose a reasonable and efficient distribution plan.Therefore,this thesis studies the optimization of drone-assisted vehicle distribution routes.Firstly,the thesis introduces customer service constraints on the basis of droneassisted vehicle distribution mode,divides customer points into three types,and establishes a mathematical model of drone-assisted vehicle distribution route optimization considering customer service constraints with the minimum transportation cost as the optimization goal.To account for the time-dependent characteristics of roads during logistics distribution,as well as the time constraints imposed by customer service windows,a time-dependent function that considers changes in vehicle speed is used.This function enables the characterization of road conditions at different times of the day.Additionally,the objective function is augmented with a penalty cost for violating time windows constraints.A mathematical model of the time-dependent drone-assisted vehicle distribution route optimization considering customer service constraints and time windows is established.Secondly,according to the characteristics of the problem,an end-to-end deep reinforcement learning framework is designed to solve the drone-assisted vehicle distribution route optimization problem.The framework can quickly solve the problem through the trained deep reinforcement learning model.Before the model training,the solution of the problem is modeled as a Markov decision process,and a deep reinforcement learning model composed of policy network and value network is built.The REINFORCE algorithm with baseline is used to train a deep reinforcement learning model,which is further accelerated by employing a node masking scheme that leverages the problem’s characteristics.During the training process,the policy network’s encoder embeds information of all nodes in the problem,and this information is then decoded by a LSTM-GRU combined network with fusion attention mechanism.The trained policy network model after training is the solution model of drone-assisted vehicle distribution route optimization problem.Finally,the training data sets and test examples of different scales are designed for the research problem,and the examples are solved by computer programming.The results of the test examples show that the deep reinforcement learning model designed in this thesis has good performance in convergence and is superior to the latest deep reinforcement learning model in convergence effect.In terms of model solving,compared with Gurobi solver,Variable Neighborhood Search algorithm and the latest deep reinforcement learning model,the deep reinforcement learning model designed in this thesis has better performance in solution quality.By setting different time-dependent and time windows parameters for comparative experiments,it is concluded that the different starting time of drone-assisted vehicle distribution considering time windows will affect the total cost of distribution.There is a certain correlation between the length of time windows and the penalty cost coefficient of time windows and the total cost of droneassisted vehicle distribution.