Actively learning dynamical systems using Bayesian neural networks

Learning dynamical systems in a sample-efficient way is important for model-based control. Active learning which sequentially selects the most informative data to sample is capable of greatly reducing sample complexity. The active learning problem for dynamical systems is hard as we can not arbitrarily draw samples from the system's state space under constraints of system dynamics. The existing approaches model the dynamical systems using Bayesian linear regression or Gaussian processes which can not be applied to complex dynamical systems with high-dimensional state spaces. In this article, we propose a new method to actively learn dynamical systems using Bayesian neural networks which allow for modeling high-dimensional systems with complex dynamics. By maximizing the accumulated differential entropies along the trajectory, the proposed method iteratively searches for the most informative action sequence which will yield informative samples when applied to the real system. With random exploration and model-based reinforcement learning as baselines, we verify the superiority of the proposed method via accuracy of one-step and multi-step predictions, the control performance, the exploration efficiency of the state space on numerical benchmarks.