Reinforcement Learning for Circular Factories

Short description

Circular factories operate under high uncertainty. Classical methods of production planning fail because of this. What is sought is a new way to reconfigure the existing production stations.

Task description

In circular factories, used products are first disassembled into their individual parts in several steps, cleaned and tested for functionality. Then the individual parts that can still be used are reassembled, with parts replaced where necessary, so that a product is created that works like new.

The simulation-based production consists of various stations, which can be modularly converted to different processes. A station can thus not only inspect individual parts, but also take on certain disassembly or assembly work. However, it has to be decided which specific process should be active at which point in time.

This task is made more difficult by uncertainties in circular factories. Today it is not known how many products will enter the factory tomorrow. The goal is therefore to develop a strategy that enables efficient production despite this uncertainty. The only room for manoeuvre lies in reconfiguring the production stations intelligently.

My work

As a student research assistant at the wbk Institute of Production Science at KIT, I work on intelligent reconfiguration management with reinforcement learning agents as part of ongoing research.

Skills

• Reinforcement Learning
• Software development
• PPO
• Graph-RL
• PyTorch
• Python