Cluster Competence in Production and Logistics: ASIM Symposium 2021
As the largest production and logistics simulation conference in Europe, the ASIM symposium presents forward-looking trends and current developments, scientific papers and interesting industrial applications every two years. From resource efficiency and digitalization topics to virtual commissioning and assistance, various aspects relevant to the use of simulation for planning, commissioning and operation of factories and logistics systems will be discussed. Marco Kemmerling from the Cybernetics Lab attended the virtual ASIM 2021 from September 15-17 and presented his work titled "Towards Production-Ready Reinforcement Learning Scheduling Agents: A Hybrid Two-Step Training Approach Based on Discrete-Event Simulations". The conference was held for the 19th time this year.© Source: Marco Kemmerling
What impressions did you personally take away from the conference?
The conference attracts researchers who work with simulations in the context of production and logistics. In particular, the topics of production planning and control and shop floor manufacturing are also represented here. After attending the conference, it can be clearly said that many researchers identify similar problems in dealing with simulations. Especially the use of data-driven models, which are a less time-consuming substitute for simulations, and their systematic derivation through data farming is a widely discussed topic.
How can your presentation topic be briefly broken down and what is its relevance to the scientific community?
In the use of reinforcement learning, one usually relies on the use of simulations. Ideally, one uses simulations that are already in use in the respective application in order to ensure an easy transfer of research into practice. Unfortunately, such simulations are often not directly suitable for the use of reinforcement learning, so that they must first be made usable by creating an interface. However, this introduces additional overhead (time) which greatly slows down the training of reinforcement learning agents. Our contribution is therefore the introduction of a multi-stage training concept, in which an agent is first trained in a self-created, tailor-made simulation, which is specifically suited for the use of reinforcement learning and represents the actual target simulation as well as possible. The majority of the training then takes place in this simulation. The agent trained in this way is then transferred to the target simulation and only needs to be briefly retrained here to compensate for any differences between the simulations. The advantage is that the amount of training in the slower target simulation is greatly reduced, but compatibility with it is still ensured.
How is your research topic embedded in the IoP complex?
The topic is addressed in the workstream B3.II in the use case "Short-term Production Planning and Control". There, reinforcement learning agents are used for planning in shop floor production. The agents can also be used in the commercial simulation software Plant Simulation, but training with this software is too time-consuming. Therefore, this is mostly carried out in a specially developed Python simulation.
How do you evaluate the implementation of such an online format in Corona times and what aspect of physical conferences do you miss the most?
The conference was organized into sessions as usual, with each session having its own breakoutroom. The sessions were well attended and there was a lively participation in the discussions following each presentation. Between the sessions there were breaks, which were conducted in Wonder, so that one had the opportunity to get into conversation there as well. To compensate for the virtual format, the conference also sent out a small care package in advance, which included key chains, snacks, etc., as well as three beers. The beers were then tried together in the evening under the guidance of a professional beer sommelier. So all in all, the organizers made every effort to make the conference as good as possible despite the circumstances. In my opinion, this also worked well, which was only possible because the participants also got involved with the format. Networking still works best in presence. That being said, in physical conferences it's easier to fully focus on the conference. In virtual conferences, you tend to do other work at the same time.