Hierarchical Bayesian Transfer Learning Between a Pair of Kalman Filters

Transfer learning strategies are typically designed in a deterministic manner, without processing uncertainty in the knowledge transfer mechanism. They also require the dependence between the participating learning procedures-Bayesian filters in this work-to be explicitly modelled. This letter develops an approach which relaxes both of these restrictive assumptions. We frame the proposed Bayesian transfer learning technique as fully probabilistic design of an unknown hierarchical probability distribution conditioned on knowledge in the form of an external probability distribution. This yields a randomized design around a base density for transfer learning which has been reported in previous work by the authors. In the Kalman filtering context, this hierarchical relaxation-which induces a knowledge-driven mixture state predictor-significantly improves tracking performance when compared to conventional transfer learning methods.