Loading [MathJax]/extensions/MathMenu.js
Multi-Agent Reinforcement Learning for Job Shop Scheduling in Flexible Manufacturing Systems | IEEE Conference Publication | IEEE Xplore

Multi-Agent Reinforcement Learning for Job Shop Scheduling in Flexible Manufacturing Systems


Abstract:

In this paper, we first outline the motivation and the need for a new approach for online scheduling in flexible manufacturing systems (FMS) based on reinforcement learni...Show More

Abstract:

In this paper, we first outline the motivation and the need for a new approach for online scheduling in flexible manufacturing systems (FMS) based on reinforcement learning (RL). We then present an initial concept for such an approach. In our method, we use Deep RL agents, who have learned to efficiently guide products through the plant and achieve near-optimal timing regarding resource allocation. Each product is controlled by its own agent, which can handle unforeseen machine failures, re-configurations of the plant topology and the consideration of local and global optimization goals. We created a virtual representation of the FMS using a Petri net, modelling the plant topology and the product flow. The agents' task is to navigate the product to the corresponding machine by choosing the according transition of the Petri net. The training consists of four stages from learning rough rules in order to fulfill a job in a Single-Agent RL setup to learning thoughtful collaboration between agents in a Multi-Agent RL (MARL) setup. We proved the feasibility of the first and second training stage by applying it to the virtual depiction of a specific research plant and obtained promising results. With this concept of a self-learning control policy, online-scheduling can be applied with less effort required to customize the setup for various FMSs.
Date of Conference: 25-27 September 2019
Date Added to IEEE Xplore: 09 March 2020
ISBN Information:
Conference Location: Laguna Hills, CA, USA

References

References is not available for this document.