Modular production control using deep reinforcement learning: proximal policy optimization

EU regulations on CO 2 limits and the trend of individualization are pushing the automotive industry towards greater flexibility and robustness in production. One approach to address these challenges is modular production, where workstations are decoupled by automated guided vehicles, requiring new...

Full description

Saved in:
Bibliographic Details
Published in:Journal of intelligent manufacturing 2021-12, Vol.32 (8), p.2335-2351
Main Authors: Mayer, Sebastian, Classen, Tobias, Endisch, Christian
Format: Article
Language:English
Subjects:
Algorithms
Automated guided vehicles
Automobile industry
Business and Management
Carbon dioxide
Control
Deep learning
Machine learning
Machines
Manufacturing
Mechatronics
Optimization
Processes
Production
Production controls
Robotics
Work stations
Workstations
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:EU regulations on CO 2 limits and the trend of individualization are pushing the automotive industry towards greater flexibility and robustness in production. One approach to address these challenges is modular production, where workstations are decoupled by automated guided vehicles, requiring new control concepts. Modular production control aims at throughput-optimal coordination of products, workstations, and vehicles. For this np-hard problem, conventional control approaches lack in computing efficiency, do not find optimal solutions, or are not generalizable. In contrast, Deep Reinforcement Learning offers powerful and generalizable algorithms, able to deal with varying environments and high complexity. One of these algorithms is Proximal Policy Optimization, which is used in this article to address modular production control. Experiments in several modular production control settings demonstrate stable, reliable, optimal, and generalizable learning behavior. The agent successfully adapts its strategies with respect to the given problem configuration. We explain how to get to this learning behavior, especially focusing on the agent’s action, state, and reward design.
ISSN:0956-5515
1572-8145
DOI:10.1007/s10845-021-01778-z