Research projects

Companies of all sizes, particularly the data-intensive Industry 4.0 sector, are confronted with increasingly complex production and logistics systems. In this context, the innovation-focused development of solutions for optimal planning and control of complex control problems is crucial to keep up with this growing complexity.

Machine learning (ML) is increasingly gaining acceptance in both research and industry to analyze and handle complex systems, as ML emulates human learning abilities using computer algorithms. In areas such as predictive maintenance, robotics, traffic control, or optimal control of transort systems, the ML sub-form reinforcement learning (RL) has been particularly effective in recent years. However, RL needs further exploration to be truly applicable to real-life challenges. The REINFORCE team will research the potential of reinforcement learning for developing optimal ways to solve complex control problems with a focus on intelligence vehicle systems.

Main research goals and use cases

REINFORCE will develop tailored methods of reinforced learning for control and planning in the field of intelligent logistics systems. A central focus of this development is to ensure that the way how automated systems learn, and which factors influence the learning process, are to be better understood by humans. The potential of reinforcement learning will be evaluated within two real-world use cases in the field of dynamic systems. Established decision-making methods already reached their limits especially due to more complex application areas.

In addition to exploring the technical aspects of reinforcement learning, such as the development of algorithms tailored to the challenges of dynamic system scenarios, the REINFORCE project also focuses on human-centered and economic aspects. Explainability of results, knowledge transfer to workers, and trustworthiness of the system and solutions are critical factors in whether the developed ML-based problem-solving agents will be adopted as viable decision-making tools in practice. The pilot project's results will be used as the basis of general guidelines for implementing reinforcement learning to tackle planning and control problems in a wide range of production and logistic systems.

Main project tasks of VRVis

Based on VRVis' extensive experience in visual analytics and explainable AI, focusing on bringing humans in the loop of data-driven decision-making processes, our researchers will mainly address the topics of explainability and user empowerment. The goal is to develop visual interfaces that enable humans to understand and visually judge how RL agents learn. At the core of these efforts lies the creation of human-centered web-based dashboards that present, among others, the decision made by the RL agent pursuing a specific strategy, a description of the trustworthiness of this action, and a summary of the repercussions of taking alternative actions.