A Hybrid Reinforcement Learning-Based Adaptive Genetic Algorithms for Order-Picking System

The Order-Picking System (OPS) is a critical determinant of warehouse efficiency, accounting for a substantial proportion of operational costs and directly influencing customer service performance. Under Industry 4.0 conditions, OPS optimization is increasingly challenged by dynamic order arrivals, time-critical fulfillment, and complex resource constraints. Conventional optimization techniques, including Genetic Algorithms (GAs), provide robust global search capabilities but suffer from limited adaptability due to static parameter configurations, often leading to premature convergence or suboptimal solutions in dynamic environments. This study proposes a hybrid Reinforcement Learning-based Adaptive Genetic Algorithm (RLAGA) that integrates an Actor–Critic reinforcement learning mechanism to dynamically adjust GA hyperparameters during the evolutionary search process. By enabling real-time adaptation of mutation probability, crossover probability, and elite selection rate, the proposed framework enhances responsiveness to evolving search landscapes in order batching and assignment problems. Extensive computational experiments conducted within a realistic OPS simulation environment demonstrate that RLAGA achieves a 1.43% reduction in total tardiness and converges 20% faster than a standard GA, albeit at the expense of increased computational time. The results confirm that reinforcement learning-driven parameter control significantly improves solution quality and convergence behavior in complex OPS scenarios. This research contributes to intelligent warehouse optimization by demonstrating the effectiveness of hybrid RL–GA architectures and provides a foundation for scalable, adaptive decision-support systems in Industry 4.0 logistics.