The IPD Institutsseminar is an ongoing course that aims to provide information about current research work at the Institute. In particular, students at the Institute are given the opportunity to report on their Bachelor's and Master's theses in front of a larger audience. The focus is on the problem definition, the solution approaches and the results achieved. However, the seminar is open to all students and employees of the KIT as well as other interested parties.
| Place | Building 50.34, Room 348 or online (see description) |
| Time | On Fridays, 10:30-12:00 / 14:00-15:30 |
The presentations must adhere to the following time frame:
- Proposal: 12 minutes speaking time + 8 minutes discussion
- Bachelor's thesis: 20 minutes speaking time + 10 minutes discussion
- Master's thesis: 30 minutes speaking time + 15 minutes discussion
If you have any questions or comments, please send an e-mail to the Institutsseminar team.
November 22, 2024 at 10:30
| Title | |
| Presenter | Marius Bohnert |
| Meeting Type | Master's thesis defense |
| Supervisor | Florian Kalinke |
| Room | Room 348 (Building 50.34) |
| Online | No link. |
| Abstract | TBD. |
November 29, 2024 at 10:30
| Title | |
| Presenter | Carolin Heidt |
| Meeting Type | Proposal |
| Supervisor | Daniel Ebi |
| Room | Room 348 (Building 50.34) |
| Online | No link. |
| Abstract | Reinforcement learning (RL) methods have become state-of-the-art for controlling agents in uncertain environments. However, the system's overall performance often depends heavily on the environment's design; even minor modifications can lead to substantially different policies. Consequently, environment shaping has gained increasing attention. Optimizing environment parameters independently, however, may result in suboptimal policies. In contrast, jointly learning the environment and control policy during the training phase offers the potential for better alignment between these components. This thesis investigates a novel environment-shaping approach that integrates the Hyperband algorithm with on-policy reinforcement learning. Hyperband, an infinite-armed bandit-based hyperparameter optimization algorithm, generates environment configurations and iteratively prunes them based on the performance of the learned policy. The proposed method is evaluated using MicroPPO, a PPO-based model for power flow management in decentralized microgrids, developed at this chair. We plan to benchmark our approach against traditional methods for parameter tuning such as Grid Search and Bayesian Optimization, as well as more recent techniques like DEPS. Future work may explore combining the principles of Hyperband and DEPS, as well as incorporating RL model hyperparameters into the optimization process to achieve further improvements. |
| Title | |
| Presenter | Hatem Nouri |
| Meeting Type | Proposal |
| Supervisor | Tobias Fuchs |
| Room | Room 348 (Building 50.34) |
| Online | No link. |
| Abstract | TBD. |
December 6, 2024 at 10:30
| Title | |
| Presenter | Jan Ettrich |
| Meeting Type | Master's thesis defense |
| Supervisor | Daniel Betsche |
| Room | Room 348 (Building 50.34) |
| Online | No link. |
| Abstract | TBD |