Dynamic crowdsourcing problem in urban–rural distribution using the learning-based approach

Inspired by real-world urban and rural distribution logistics scenarios, this study explores the dynamic crowdsourcing multi-depot pickup and delivery problem (DCMDPDP) through an online platform (OCP), where requests and crowdsourced vehicles arrive dynamically. Vehicles either collect from multiple depots for deliveries or pick up from customers to depots. To maximize the OCP’s daily total gain, the net value of completed task revenue minus vehicle compensation costs, we integrate anticipated future gains into each decision-making process and formulate the DCMDPDP as a Markov decision process. A learning-based hybrid heuristic algorithm is proposed for the DCMDPDP. Specifically, we develop an enhanced adaptive large neighborhood search algorithm leveraging the heat map to batch orders into multiple groups and assign them to depots, where the heat map is learned offline using a graph convolutional residual network with an attention mechanism model. A value learning-based algorithm is also developed to obtain optimal matches between order batches and vehicles, and near-optimal travel routes. Experimental results demonstrate that the proposed algorithm improves the OCP total gain by 46.09%, 57.13%, 0.49%, 2.45%, 1.08%, and 2.77% over six benchmarks. Furthermore, the proposed algorithm reduces unserved customers to 7.83 on average, outperforming six benchmarks by 2.19–167.52 fewer cases. Moreover, extensive experiments validate that the proposed algorithm is strongly generalizable in handling instances with varying customer sizes and different temporal, spatial, and demand distributions.