Dr. Kirill Djebko

Kirill Djebko has been working at the University of Würzburg since 2015 and completed his doctorate in 2020. He worked on the KINERGY project from 2019 to 2024. Between 2022 and 2023 he was involved in the VeriKI project and since 2024 he has been working on the LeLaR project.

Research Interests include but are not limited to: Metaheuristics in simulation with a focus on genetic algorithms, optimization and regression algorithms as well as deep reinforcement learning for practical applications.

KINERGY - Intelligent heating system monitoring and optimization

The goal of the KINERGY project is to optimize heating systems. These optimizations are performed by first developing a digital model of a real heating system with parameters corresponding to settings and internal parameters, and then calibrating these internal parameters with measured data from the real system, while leaving the adjustable settings parameters unchanged. Finally, the calibrated model is used to calculate behavioral changes that occur when the settings parameters are changed while the internal parameters remain fixed. The modified values for the settings parameters are provided by a Decision Support System (DSS), which is being developed as an additional component within the KINERGY project. The DSS in turn receives the state of the real heating system from a third component, Deadalus, a web agent that analyzes measured data and periodically calls the DSS to either detect malfunctions or check for a potential efficiency improvement. Both can lead to  changes of the settings parameters. These settings parameters are then passed to the simulation component, a simulation is performed, and the resulting simulated data is analyzed to see if the change has resulted in an improvement. In a later step, the improved settings parameters are to be applied to the physical heating system in order to evaluate its effect on the real heating system.

VeriKI - Verification of AI methods in decentralized structures for New Space applications 

As part of the VeriKI project, an AI-based position controller that is robust against variations in moments of inertia is to be developed. For this purpose, Deep Reinforcement Learning is used to train an AI agent through autonomous interaction with a simulation environment. The feasibility of this approach to develop a high-quality and robust AI-based attitude controller that can be used without manual fine-tuning is demonstrated.

LeLaR - In-Orbit Demonstrator Learning Attitude Control

The goal of the LeLaR project is to train a complete AI attitude controller for the InnoCube nanosatellite using Deep Reinforcement Learning and to evaluate the resulting controller in orbit. The AI attitude controller should be robust against natural and unexpected deviations from the expected operating conditions. It will be trained on the ground in simulation and then uploaded to the satellite. The project aims to demonstrate that AI attitude controllers can achieve sufficient quality to be used in real missions while retaining their robustness properties. In particular, it should be shown that a sim-to-real transfer is possible in this form.