Theory and experimental analysis of cognitive processes in early learning

Abstract

This paper presents an algorithm of unsupervised learning for applications in robotics and a knowledge structure which supports the behaviour generation (BG) module of the RCS/NASREM architecture designed at NIST. Minimum initial knowledge is presumed ("bootstrap knowledge"). The learning system uses the newly arrived information to extract rules of motion and construct a multiresolutional world model (WM). It evolves as a structure of knowledge representation which allows the BG to create and execute plans at each level of resolution. The concept of recursive generalization is explored as the main tool of rule extraction and knowledge organization. The experiment in learning is described based upon simulation of a 2D and a 3D mobile system.