Competitive Learning Algorithms in Adaptive Educational Toys

Unsupervised neural learning is typically employed in dimensionality reduction, to extract relevant features for subsequent stages of supervised learning. In this paper we examine a class of unsupervised learning algorithms used for a somewhat different purpose, that of clustering input vectors into various learned stereotyped behaviours in mobile robots [1] . Unsupervised techniques have significant advantages in such applications given their ability to continuously and autonomously adapt to a changing operating environment. While supervised algorithms require a set of labelled training data, the unsupervised techniques evolve input-output mappings based entirely on the input data which may be acquired from the environment in real-time.