Quantifying the Stand-alone Benefit of Dueling DQN for Job-Shop Scheduling: A Systematic Significance Study. 

Bian Yusong

Abstract

The Job Shop Scheduling Problem (JSP) is a classical optimization problem, which aims to process multiple jobs consisting of several processes under limited machine resources and it often has a specific order. The key to solve this problem is to rationalize the order and timing of processing (makespan). This study investigates the impact of Dueling structure in deep reinforcement learning on job shop scheduling performance. In our experiments, we compare the performance between DDQN and Dueling-DDQN, by keeping the same training framework and hyperparameter configurations. Research excludes the influence of other variables and focus on the contribution of the network structure itself. To ensure the validity and reproducibility of the results, this study was conducted for 1500 episodes with fixed cycles of unexplored validation. Research focuses on evaluating whether the Dueling structure can significantly reduce the average makespan of job shop problem, providing evidence of its reliability in specific instances and laying the groundwork for future experiments scaled up to larger datasets.