A Reinforcement Learning-Based Two-Stage Memetic Algorithm for Event-Driven Multi-Compartment AGV Rescheduling Problem

Automated Guided Vehicles (AGVs) are extensively utilized in logistics and transportation systems. However, improving transport efficiency remains challenging due to dynamic events such as urgent tasks, task cancellations, and AGV failures. To address this, a mixed integer linear programming model is first formulated. A reinforcement learning based two-stage memetic algorithm (RTMA) is proposed for the event-driven AGV rescheduling problem. In RTMA, the first phase is the normal scheduling phase, and the second phase is the event-driven rescheduling phase. An emergency response mechanism is designed to bridge these two phases, enabling rapid schedule update and resource reallocation. A novel heuristic algorithm is developed to generate a high-quality initial solution by exploiting task distributions. To further enhance resource utilization, a delayed waiting strategy for multi-compartment AGVs is introduced. Then, reinforcement learning is embedded into the local search process to select problem-specific neighborhood structures, with a noise attenuation strategy dynamically balancing exploration and exploitation. Furthermore, three theorems are formulated and proven to improve the efficiency of feasibility checks. The proposed RTMA is evaluated using real-world production data from an electronics manufacturing workshop. Experimental results demonstrate the superiority of RTMA in handling dynamic events and optimizing transport efficiency compared to existing advanced methods.