Active Learning in the Sensorimotor Loop

Martius, Georg
In this thesis we study a novel approach to on-line learning of artificial neural networks, called backward modelling, and apply it to active learning in the sensorimotor loop. At first the mathematic foundations of this approach are elaborated. We observe effects like spontaneous symmetry breaking, response increasing, and generalisation improvement at a theoretical level. We then justify the theory with experimental results on some synthetic problems, in order to understand the phenomena clearly. Finally we consider a simple robot with an adaptive world model. In the case the controller of the robot is just covering a sub-space of the actuator space we realise degenerated world representations in the world model with passive learning and standard learning algorithms. We show that backward modelling and active learning point out degeneracies in the world model and correct them with direct exploration. A special kind of active learning evolves from the use of backward modelling which directly queries patterns on the fly. Additionally, different strategies are investigated in order to control the interplay of controller based and active learning based behaviour.
Pubdate / Erscheinungsdatum: 
Pages / Seitenanzahl: 
2005-4.pdf1.16 MB