Multi-agent reinforcement learning for order dispatching and sequencing in on-demand food delivery platforms

Online on-demand food delivery platforms have substantially transformed the way millions of people live worldwide, providing convenience to users while creating significant economic value and employment opportunities. These platforms stand out for their scale and operational complexity. Despite the abundance of historical and real-time data available, optimizing such systems remains a challenging problem that involves dynamic resource allocation, stochastic demand, and the coordination of large fleets of drivers to fulfill delivery tasks. This thesis tackles the order dispatching and delivery sequencing problem using reinforcement learning (RL) within a simulated environment calibrated with real-world data. The delivery process is formulated as a sequential decision-making problem, in which RL agents learn to improve service quality by minimizing average delivery time and delivery-time variance. We begin with a single-agent setting and subsequently extend it to a multi-agent scenario, where multiple drivers operate concurrently under partial observability. A reward function is designed to guide agent learning. To enable coordination and stable policy updates, a Centralized Training with Decentralized Execution (CTDE) architecture with policy sharing is employed for the multi-agent setting, allowing agents to learn joint value functions during training while acting independently during execution. In addition, a communication channel between agents is introduced to facilitate coordination. Experimental results show that the RL agents learn effective and cooperative delivery strategies, outperforming baselines in both single- and multi-agent scenarios. These findings demonstrate the potential of RL-based approaches to enhance decision-making in real-time logistics systems.