Active Learning in Motor Control
Philipp Robbel, Marc Toussaint, Sethu Vijayakumar · 2007
In this dissertation we consider how the performance of a learning control system can be improved with an efficient exploration strategy. We apply principled approaches from active learning to realise task-specific exploration with the LWPR on-line learning scheme. Our algorithm is based on the confidence in the model predictions and directs exploration to areas of high uncertainty. To the best of our knowledge, this is the first clear strategy for active learning in low-level motor control scenarios. Using two simulations, we show that learning the inverse dynamics of a movement system can benefit from an active data selection strategy. Both simulations – a Newtonian particle and a compliant two-joint robot arm – also provide an intuitive real-time visualisation of the LWPR confidence bounds and the explored space. The suggested algorithm is shown to be superior to simpler exploration schemes such as random flailing of the robot arm.