National University of Singapore
Suzhou Research Institute
311 Program Final Year Project (2024/2025)
National University of Singapore
Suzhou Research Institute
311 Program Final Year Project (2024/2025)
The Vehicle Routing Problem (VRP) remains a critical challenge in logistics, requiring efficient solutions to minimize costs and environmental impact. Traditional optimization methods struggle with scalability and dynamic constraints, prompting the exploration of deep reinforcement learning (DRL) as a transformative approach. This study adapts a classical Capacitated VRP (CVRP) model to a real-world e-commerce logistics data set, integrating domain-specific constraints to enhance the practicality of DRL-based solutions. We propose a residual edge-graph attention network combined with Proximal Policy Optimization (PPO) to address routing challenges, benchmarking against the Gurobi solver. Results demonstrate that the DRL model achieves competitive solution quality with significantly reduced computational latency (e.g., 3 seconds vs. 600 seconds for 50-node instances), highlighting its potential for real-time, large-scale logistics optimization. Challenges in generalization and input flexibility persist, but hybrid models and algorithmic advancements offer promising avenues for future research.