Noisy-or nodes for conditioning models

Abstract

In order to adapt the behavior of robots to varying environments, conditioning models provide interesting ideas. A prediction system is an important part of such models. The problem is to update it according to the sequence of stimuli perceived by the robot. Bayesian networks can be used to implement the prediction system. However, update rules are very complex and we need an incremental and fast learning process. We propose the use of noisy or nodes with appropriate learning rules. Numerous features of conditioning have been tested and promising results have been obtained. © 2010 Springer-Verlag.