A QMIX-Based Multi-Agent Reinforcement Learning Approach for Crowdsourced Order Assignment in Fresh Food Retailing

Crowdsourced delivery plays a key role in fresh food retailing, where tight time limits and perishability require fast, reliable fulfillment. However, real-time order–courier assignment is challenging because orders arrive in bursts, couriers’ locations and availability change, capacities are limited, and many decisions must be made simultaneously. We propose Attn-QMIX, a novel attention-augmented multi-agent reinforcement learning framework that models each order as an agent and learns coordinated matching strategies through centralized training with decentralized execution. The framework develops a new capacity-aware multi-head attention mechanism that captures complex order–courier interactions and dynamically prevents courier overload and integrates it with a QMIX-based mixing network equipped with hypernetworks to enable effective credit assignment and global coordination. Extensive experiments on a real-world road network show that Attn-QMIX outperforms five representative methods. Compared with a novel cooperative ant colony optimization method, it reduces total cost by up to 2.30% while being up to 3403 times faster in computation.