Solving industrial chain job scheduling problems through a deep reinforcement learning method with decay strategy

Abstract

A Job Shop Scheduling Problem (JSSP) is an NP-Hard problem with extensive applications in many domains such as transportation and manufacturing industrial chains. Deep Reinforcement Learning (DRL) has emerged as a novel approach being distinct from traditional scheduling and heuristic methods. Although DRL has shown promising results in addressing JSSP, several limitations remain, such as ignoring an optimal solution space and lacking the focus in policy network learning, which affects both scheduling quality and learning speed. To address these issues, we introduce the DecayP30 method that incorporates the solution space partition and dynamic weight allocation into the decision-making process. Specifically, the DecayP30 method innovatively replaces the traditional clipping operation in Proximal Policy Optimization (PPO) with the Sigmoid function, a key feature of the Preconditioner Proximal Policy Optimization (P30) approach. We introduce a dynamic decay strategy to address the “heavy-head and light-tail” issue JSSP. The new approach ensures a more comprehensive solution space while emphasizing sequential relationships inherent in JSSP. We evaluate the new method in four major JSSP datasets. Extensive experiments demonstrate that our proposed method exhibits better convergence speed and scheduling quality compared to most the DRL methods.