Deep Reinforcement Learning for Route Optimization in E-commerce Return Management

Abstract

E-commerce return volumes continue to surge, presenting significant operational challenges for reverse logistics management. Traditional route optimization approaches struggle with the dynamic nature of return collection requests characterized by unpredictable arrivals, stringent time windows, and dispersed locations. This paper proposes a deep reinforcement learning framework combining graph attention networks with proximal policy optimization for dynamic return routing. The approach encodes spatial relationships between pickup locations and learns adaptive policies through continuous environmental interaction. Experimental evaluation on synthetic benchmarks and real-world data demonstrates substantial improvements: 15.8% reduction in routing distance and 11.4 percentage point improvement in on-time pickup rates compared to genetic algorithm baselines. Computational analysis shows 0.28-second inference times enabling real-time adaptation. Results validate practical viability for intelligent reverse logistics optimization.