Abstract

Model based predictions of future trajectories of a dynamical system oftensuffer from inaccuracies, forcing model based control algorithms to re-planoften, thus being computationally expensive, suboptimal and not reliable. Inthis work, we propose a model agnostic method for estimating the uncertainty ofa model?s predictions based on reconstruction error, using it in control andexploration. As our experiments show, this uncertainty estimation can be usedto improve control performance on a wide variety of environments by choosingpredictions of which the model is confident. It can also be used for activelearning to explore more efficiently the environment by planning fortrajectories with high uncertainty, allowing faster model learning.

 

Quick Read (beta)