Proposal of an Intrinsically Motivated System for Exploration of Sensorimotor State Spaces
Matthias Kubisch, Manfred Hild, Sebastian Höfer · 2010
For a well-adapted behavior, an individual has to establish and refine its own body model continuously during life time. The best way to collect the necessary data for bootstrapping this model is active movement. At this, it is useful if actions are chosen in such a way that the gathered information matches well with the current stage of the learning system. In this paper, we investigate the method of self-exploration by intrinsic motivation, whereby the individual is driven to select appropriate actions to support its own learning progress. We implemented an unsupervised neural multi-expert architecture and tested the learning algorithm on an abstract artificial individual.