Bioinspired dynamic model adaption for AUVs

Autonomous Underwater Vehicles (AUVs) have received a lot of attention in recent years. To operate independently they have to be able to maneuver securely in often changing environments with changing viscosity and currents or even biofouling and corroded or damaged hulls.
Unlike the usually handcrafted and thus static models most commonly used in AUVs today, biological systems including humans use adaptable models that are constantly checked for their validity.
Based on existing works using support vector regression, this thesis will focus on implementing an algorithm to mimic this capability, employing deep neural networks with dropout as a bayesian approximation serving as a measure of uncertainty to gauge the validity of the current model.
The algorithm will be evaluated on existing data obtained from Dagon and compared to the existing algorithm.