Deep reinforcement learning algorithm incorporating problem characteristics for dynamic multi-objective permutation flow-shop scheduling problem

doi:10.1016/j.swevo.2025.101973

Deep reinforcement learning algorithm incorporating problem characteristics for dynamic multi-objective permutation flow-shop scheduling problem

2025-07-01

0

PRE

AI

DOI：10.1016/j.swevo.2025.101973

Original

Original request for help

Save

Abstract

En 中文

Dynamic permutation flow shop scheduling problem (DPFSP) plays a critical role in real-world production systems, characterized by complex uncertainties including machine breakdowns, variable processing times, and the unpredictable job arrivals. Developing real-time solution approaches for such complex dynamic environments represents both a significant industrial need and a substantial computational challenge. Deep reinforcement learning (DRL) has demonstrated excellent capability for rapid and adaptive decision-making in complex dynamic environments, making it particularly suitable for DPFSP applications. This paper proposes a novel DRL algorithm incorporating problem characteristics (NDRLA_IPC) that specifically addresses the dynamic multiobjective PFSP (DMPFSP), with the objectives of minimizing the weighted maximum completion time and the total tardiness. NDRLA_IPC leverages a key insight that DMPFSP can be decomposed into a series of static PFSPs requiring real-time solutions, and implements a double deep Q-network (DDQN) architecture with components specifically engineered for DMPFSP characteristics. The algorithm introduces three key innovations: (1) a state feature vector design with high discrimination and generalization capabilities; (2) an action space designed to minimize temporal gaps between operations on the Gantt chart by leveraging processing constraints dynamically derived from the evolving problem state; and (3) a theoretically-validated reward function that effectively evaluates the online execution impact of each action. Comprehensive experiments demonstrate that NDRLA_IPC, after training on small-scale instances, transfers effectively to larger-scale DMPFSPs, delivering high-quality realtime solutions that outperform existing approaches across multiple performance metrics.

Keywords:

Deep reinforcement learning

Dynamic permutation flow shop scheduling

problem

Real-time scheduling

Multi-objective optimization

Deep reinforcement learning algorithm incorporating problem characteristics for dynamic multi-objective permutation flow-shop scheduling problem

Abstract

Journal

Organization