A reinforcement learning model for dynamic distribution route optimization of agricultural products under replenishment demand

Most existing logistics optimization models assume static demand or ignore transportation replenishment, which limits their applicability in dynamic agricultural environments. Therefore, this paper constructs a dual-objective multi-center cold chain distribution model to minimize total cost and maximize customer satisfaction. It incorporates dynamic customers and in-transit replenishment into the constraints and introduces a Reinforcement Learning-guided Adaptive Large Neighborhood Search (RL-ALNS) algorithm for the solution. Actor-Critic deep reinforcement learning designs state-action spaces and reward functions for dynamic delivery environments. It iteratively updates paths by jointly training strategy networks and value networks. Then, it embeds reinforcement learning-guided adaptive large neighborhood search into the solution process, dynamically selecting destruction-repair operators to improve convergence efficiency and stability in large-scale, real-time updating scenarios. Experimental results demonstrate that, under the same customer scale, allowing vehicles to restock en-route can reduce total costs from 2920.12 RMB to 2246.32 RMB, while improving customer satisfaction from 0.8613 to 0.9346. For a medium-sized instance with 52 nodes, the optimal cost achieved after five independent runs of the algorithm is 2246.64 RMB. In small-scale benchmark tests, the RL-ALNS reduces the solution time from 28 s to 7.04 s.