The
economic and ecological added value for manufacturing companies crucially
depends on how efficiently they handle dynamic events in production and conduct
real-time scheduling. Individual customer orders present a significant
challenge for overall planning. Therefore, both reactive and proactive planning
strategies are necessary to increase the agility and robustness of the
production system. Rule-based heuristics and metaheuristic algorithms are often
used in complex planning tasks. While heuristics provide quick solutions, their
rigidity often affects quality. Metaheuristic algorithms iterate various
approaches to improve optimization but reach their limits with increasing
complexity and often get trapped in local optima, reducing efficiency. Reinforcement
Learning (RL) offers a promising alternative here. RL continuously learns from
the environment and adapts its strategies, enabling production systems to
respond to unexpected events and improve their performance. This leads to more
robust and agile planning that is better aligned with the challenges of modern
production paradigms.
In this context, the following research topics are
offered as part of the work:
- Literature review on relevant fundamentals and
suitable methodological approaches, including:
- Fundamentals of (Deep) Reinforcement Learning
(DRL)
- Fundamentals of agent-based process chain
simulation
- AI-based
dynamic
production
scheduling
- Types of shop-floor production systems (e.g.,
flow shop, job shop, etc.) relevant to industries like battery manufacturing,
surface technology, semiconductor production, and their respective
process chains
- Modeling and simulating production process chains
as DRL training environments using open-source tools such as AnyLogic, SimPy,
OpenAI Gym etc.
- Defining appropriate state space, action space,
and reward function for DRL-based dynamic scheduling (e.g., make-span,
machine utilization, energy consumption, total tardiness, etc.)
- Connection and Interaction between production
chain simulation models and DRL training algorithms
- Selection of suitable DRL algorithms for training
and validation (e.g., DQN, policy-based algorithms, etc.)
- Evaluation of productivity, economic, and
ecological outcomes based on the defined reward function
- Critical
evaluation and outlook
Depending
on the scope of the work, only some of the above points may be covered. Experience with programming languages such as
Python and Java is advantageous but not mandatory. If you are
interested in these topics, I would be pleased to receive your email with the
corresponding documents at chao.zhang@tu-braunschweig.de.
Art:
- Studienarbeit(Master)
- Masterarbeit
Fachrichtungen: Computer science, mechanical engineering, industrial engineering, electromobility, etc.
Begin der Arbeit: Sofort
Zuletzt geändert: 15.08.2024