Deep reinforcement learning for dynamic vehicle routing with demand and traffic uncertainty

The capacitated vehicle routing problem with dynamic demand and traffic conditions presents significant challenges in logistics and supply chain optimization. Traditional methods often fail to adapt to real-time uncertainties in customer demand and traffic patterns or scale to large problem instances. In this work, we propose a deep reinforcement learning framework to learn adaptive routing policies for dynamic capacitated vehicle routing problem environments with stochastic demand and traffic. Our approach integrates graph neural networks to encode spatial problem structure and proximal policy optimization to train robust policies under both demand and traffic uncertainty. Experiments on synthetic grid-based routing environments show that our method outperforms classical heuristics and greedy baselines in minimizing travel cost while maintaining feasibility. The learned policies generalize to unseen demand and traffic scenarios and scale to larger graphs than those seen during training. Our results highlight the potential of deep reinforcement learning for real-world dynamic routing problems where both demand and traffic evolve unpredictably.