Actively Learning Gaussian Process Dynamical Systems through Global and Local Explorations

Usually learning dynamical systems by data-driven methods requires large amount of training data, which may be time consuming and expensive. Active learning, aims at choosing the most informative samples to make more efficient is a promising way solve this issue. However, actively difficult since it not possible arbitrarily sample state-action space under constraint system dynamics. The state-of-the-art for iteratively search an pair maximizing differential entropy predictive distribution, or long trajectory sum variances along trajectory. These suffer from low efficiency high computational complexity memory demand. To these problems, paper proposes novel sample-efficient combine global local explorations. As exploration, agent searches relatively short in whole system. Then, as action sequence optimized drive system’s state towards initial found exploration explores Compared methods, proposed are capable exploring efficiently, have much lower With baselines, advantages verified via various numerical examples.