@phdthesis {RolfThesis2012,
	title = {Goal Babbling for an Ecient Bootstrapping of Inverse Models in High Dimensions},
	author = {Matthias Rolf},
	year = {2012},
	abstract = {Learning to coordinate high-dimensional motor systems is a fundamental task for humans as well as robots. Traditional approaches to the computational learning of coordination skills rely on an exhaustive exploration of possible actions, which is not feasible in high dimensions. This thesis investigates reaching as a prototypical coordination problem and introduces the concept of goal babbling for the learning of reaching skills in high-dimensional domains. Goal babbling is inspired by studies about infant development that show that already newborns attempt goal-directed movements, even if they can not perform them successfully. This thesis develops methods that bootstrap reaching skills by mimicking such early goal-directed movements, and demonstrates their success in high dimensions. The methods developed in this thesis implement goal babbling for the learning of inverse models as a direct mean to solve coordination problems. Theoretical results show how such inverse models can be learned by means of goal babbling. This thesis introduces the first algorithm that can learn inverse models by fitting observed data even when the coordination problem contains solution sets that are not convex, which has been a severe limitation of previous algorithms. It is shown that the approach allows for a bootstrapping that scales almost constantly with respect to the dimension of the action space, which is opposed to the exponential cost of exhaustive exploration. Experiments demonstrate that goal babbling constitutes a positive feedback loop between exploration and learning during the initial bootstrapping of skills. In an online learning scenario this is shown to permit substantial speedups of learning and to allow for human-level learning speed. Reaching with a bionic robot trunk is investigated as a practical scenario that is very hard to solve without learning due to the lack of analytical models and non-stationary system behavior. Extensive real-world experiments demonstrate the practical feasibility and usefulness of the goal babbling approach on this challenging platform.},
	institution = {University of Bielefeld}
}