A PSO-Assisted Reinforcement Learning Algorithm for Job Shop Scheduling

The Job Shop Scheduling Problem (JSSP) is a critical and complex optimization challenge in manufacturing. Recently, Reinforcement Learning (RL)-based methods have garnered significant attention due to their promising performance in improving scheduling efficiency. However, traditional RL algorithms often suffer from inefficiencies in exploring the vast solution space, resulting in suboptimal policies. This paper introduces a novel Reinforcement Learning algorithm assisted by particle swarm optimization (PSO), called PSO-RL, to address these limitations. We design a multiple-particle searching framework for the RL algorithm, where multiple solutions can be improved synchronously by an RL model. During the training process, PSO is periodically applied to guide the worst solutions towards the global and local optima identified so far. By integrating PSO's global search capability with RL's adaptive learning, our PSO-RL algorithm achieves a balanced exploration-exploitation tradeoff. Experimental results on benchmark JSSP instances demonstrate that PSO-RL outperforms state-of-the-art methods, yielding lower makespans. The adaptability of the PSO-RL framework makes it a versatile tool for various industrial applications, enhancing the performance of production planning and intelligent manufacturing systems.