Adaptive autonomous control using online value iteration with gaussian processes

摘要

In this paper, we present a novel approach to controlling a robotic system online from scratch based on the reinforcement learning principle. In contrast to other approaches, our method learns the system dynamics and the value function separately, which permits to identify the individual characteristics and is, therefore, easily adaptable to changing conditions. The major problem in the context of learning control policies lies in high-dimensional state and action spaces, that needs to be explored in order to identify the optimal policy. In this paper, we propose an approach that learns the system dynamics and the value function in an alternating fashion based on Gaussian process models. Additionally, to reduce computation time and to make the system applicable to online learning, we present an efficient sparsification method. In experiments carried out with a real miniature blimp we demonstrate that our approach can learn height control online. Further results obtained with an inverted pendulum show that our method requires less data to achieve the same performance as an off-line learning approach.