A revised deep reinforcement learning algorithm for parallel machine scheduling problem under multi-scenario due date constraints

doi:10.1016/j.swevo.2024.101808

2025-02-01

PRE

Amir M. Fathollahi‐Fard *

Guangdong Tian

DOI：10.1016/j.swevo.2024.101808

原文链接

原文求助

摘要

En 中文

Efficient scheduling in storage chip manufacturing is crucial for meeting market demands and maintaining competitiveness in the semiconductor industry. This paper explores the complex parallel machine scheduling problem in this context, with a particular focus on scenarios involving both strong and weak due date constraints. The primary objective is to minimize the combined total of weighted tardiness and completion time, essential for enhancing production efficiency and reducing costs. To address this, we develop a mixed-integer programming (MIP) model and rigorously prove the NP-hardness of the problem, emphasizing its computational complexity and the challenges of finding optimal solutions. Given these challenges, we propose a revised Deep Reinforcement Learning (DRL) framework based on the Deep Q-Network (DQN) to effectively tackle the scheduling problem. The scheduling of both sampling test and ordinary wafers is modeled as a Markov Decision Process (MDP), allowing the application of advanced reinforcement learning techniques. To further improve learning efficiency and stability, we employ the Double DQN (DDQN) algorithm, which mitigates the overestimation bias often encountered with standard DQN approaches. We conduct extensive comparative experiments, evaluating the performance of the DDQN algorithm against other DRL, heuristic, and metaheuristic algorithms across diverse scheduling scenarios. Robust statistical analyses, including the Friedman test and Nemenyi post-hoc test, are used to rigorously compare performance across competing methods. The results consistently show that the DDQN algorithm outperforms alternative methods, demonstrating its potential as a powerful tool for optimizing complex scheduling tasks in storage chip manufacturing.

Keyword:

Due date constraints

Deep reinforcement learning

Deep Q-Network

Markov Decision Process