1511. Direct action selection for solving the dynamic job shop scheduling problem with deep reinforcement learning
Invited abstract in session MD-38: (Deep) Reinforcement Learning for Combinatorial Optimization, stream Data Science meets Optimization.
Monday, 14:30-16:00Room: Michael Sadler LG19
Authors (first author is the speaker)
| 1. | Fynn Martin Gilbert
|
| Decision and Operation Technologies, Bielefeld University | |
| 2. | Kevin Tierney
|
| Business Decisions and Analytics, University of Vienna |
Abstract
This study compares hyperheuristic and direct job selection strategies in deep reinforcement learning (DRL) for the dynamic job shop scheduling problem (JSSP). Both hyperheuristics and direct job selection have emerged as prominent action selection mechanisms in DRL for solving JSSPs, particularly in dynamic, uncertain environments that include machine breakdowns or variable processing durations, which better reflect real-world manufacturing challenges. While direct job selection and especially hyperheuristic strategies are widely used in DRL for job shop scheduling, their relative strengths and limitations remain an open question. To explore its relative strengths, we investigate direct job selection as an alternative to hyperheuristic agents for makespan minimization by comparing their performance, training efficiency, and generalizability. We conduct experiments in a discrete event simulation environment, training both DRL approaches across different problem sizes and degrees of variability and flexibility. This comparison highlights the critical role of action selection mechanisms in DRL-based solution procedures for the dynamic JSSP and provides empirical guidelines for designing effective DRL scheduling policies.
Keywords
- Scheduling
- Machine Learning
- Combinatorial Optimization
Status: accepted
Back to the list of papers