Research on Intelligent Optimization Algorithm for Multi-Objective Vehicle Routing Based on Deep Reinforcement Learning

Abstract

To address the problems of slow convergence and poor multi-objective coordination in complex scenarios of traditional multi-objective vehicle routing optimization algorithms, as well as the inaccurate state representation and policy update drift of existing deep reinforcement learning algorithms, this paper proposes an enhanced dynamic self-attention and adaptive reward-based PPO algorithm (ADAR-PPO) to solve the multi-objective vehicle routing optimization problem. First, a mathematical model for multi-objective vehicle routing optimization is constructed, clarifying optimization objectives such as shortest path, lowest transportation cost, and load balancing, along with constraints such as vehicle capacity and time windows. Second, an enhanced dynamic self-attention encoder is designed, integrating channel attention and SwiGLU activation units to achieve dynamic updating and deep fusion of node features, improving the accuracy of state representation. Then, an adaptive multi-objective reward mechanism is proposed, dynamically adjusting the weights of each objective and generating a comprehensive reward by combining constraint penalties to guide the agent in learning the Pareto optimal solution. Finally, the PPO policy update method is improved by introducing advantage normalization and dynamic pruning thresholds to enhance the algorithm's training stability and sample efficiency. Experimental results based on the Solomon standard dataset show that, in the C101 instance, the ADAR-PPO algorithm achieves an HV value of 0.896, a 12.8% improvement over the standard PPO and a 23.5% improvement over GA; an IGD value of 0.087, a 49.4% reduction over the standard PPO and a 61.3% reduction over PSO; an average path length of 892.3 m, a 12.6% reduction over GA; an average transportation cost of 702.5 yuan, a 7.6% reduction over PSO; and a convergence iteration count of 180, a 30% reduction over the standard PPO. These results demonstrate that the proposed algorithm outperforms the comparative algorithms in multi-objective collaborative optimization, convergence speed, solution accuracy, and perturbation adaptability, providing effective technical support for intelligent planning of logistics delivery routes.