CSAIL Event Calendar: Previous Series
Bayesian Inference for Efficient Motor Learning
Speaker: Marc Deisenroth , Cambridge
Abstract: Learning from experience is a key ingredient in the behavior of intelligent beings and holds great potential for artificial systems. Humans and animals use experience to learn complicated tasks relatively quickly, that is, they do not require many trials to succeed. In contrast, artificial learners cannot generally replicate this speed of learning. To speed up artificial learning, we borrow two key ingredients that make biological learning so successful: the ability to generalize and the explicit incorporation of uncertainty into the decision-making process. We use probabilistic Gaussian process models for predictions to explicitly account for both key ingredients. We successfully apply our learning algorithm to control dynamic systems with continuous state and action spaces. Our policy search algorithm learns to solve complicated tasks, such as the cart-pole swing up, the Pendubot, or the cart-double pendulum swing up, in a couple of trials. The algorithm works in simulation and in hardware, and we report an unprecedented speed of learning for the above-mentioned control problems.