Deep Reinforcement Learning for Setup and Tardiness Minimization in Parallel Machine Scheduling

This paper introduces a novel deep reinforcement learning algorithm for the identical parallel machine scheduling problem, with a focus on minimizing setup time and tardiness. In modern manufacturing environments, accommodating diverse consumer demands has emerged as a primary challenge, shifting focus towards small-batch, multi-product production schemes. However, traditional optimization techniques pose challenges due to the inherent complexity of production environments and uncertainties. Recently, reinforcement learning has emerged as a promising alternative for scheduling in such dynamic environments. We formulate the scheduling problem as a Markov decision process, with states designed to capture the key performance indicators related to setup and tardiness. The algorithm's actions correspond to selecting one heuristic rules among SSPT, ATCS, MDD, or COVERT. During the learning phase, we employ the Proximal Policy Optimization (PPO) algorithm. Experimental results on a custom dataset demonstrate superior performance compared to individual applications of existing heuristics.