Tobias was working on "The Role of Temperature-Induced Stochastic Noise in Reinforcement Learning-Driven Microswimmer Control" supervised by Samuel Tovey. We congratulate him on this wonderful achievement and invite you to attend the defense.
Abstract: This work aims at getting insight into how reinforcement learning controlled microswimmers adapt to temperature changes in their deployment environment through computer-based simulations. And to develop a training strategy, that can be applied for further investigations,when studying Brownian Motion. As the benchmark task, the agents were trained at constant temperatures to rotate a rod. The resulting neural networks were later deployed under varying conditions, while the success of the emerged strategies was measured by rotational velocity. In conclusion, high temperature trained neural networks surpassed the abilities of those,trained at a lower temperature. When controlling microswimmers with the latter ones, the rod’s rotational velocity quickly dropped off with increasing temperature, whereas higher-temperature trained microswimmers manage to maintain a constant velocity over the entire defined temperature span. Therefore, when using reinforcement learning controlled microswimmers in fluctuating temperature environments, it is sufficient to train them only atthe highest expected temperature. Thus, a viable training strategy was found.
Best of luck Tobias on your upcoming defense and future endeavors!
