Occasional drivers’ allocation using two-stage deep reinforcement learning – a case study of an e-grocery platform

This study addresses a stochastic crowd-shipping problem, considering multiple customer types and the uncertain availability of occasional drivers. The objective is to minimize the total delivery cost through efficient customer allocation and routing for dedicated and occasional drivers. A novel two-phase Deep Reinforcement Learning approach is introduced to efficiently handle large-scale instances. In the first stage, a Single-Layer Feed-Forward Neural Network is implemented to train and validate the estimated reward function of random sequences using an on-policy method. The customer types are classified by taking appropriate actions using a Mixed-Integer Programming solution. Then, a modified Proximal Policy Optimization algorithm updates policies over multiple epochs during training. After allocation, optimal routes for occasional drivers are determined by a Capacitated Vehicle Routing model. A case study of an online e-grocery platform illustrates the efficiency of the proposed system. Experimental results indicate that the proposed approach outperforms the previous approach by approximately 17%.