Competitive learning algorithms in adaptive educational toys

We examine a class of unsupervised learning algorithms used for clustering input vectors into various learned stereotyped behaviours in mobile robots. We examine the application of one class of unsupervised algorithms known as competitive learning (CL) and investigate its suitability as an adaptive control mechanism for an educational toy. Two variants of competitive learning, hard competitive learning and soft competitive learning, are explored. These explorations take into account limitations of the robotic system which restrict the complexity of the algorithm which can be realized. Children develop a great deal of their understanding of the world through play and this process can be enriched by providing a child with a stimulating toy. Anyone who has spent time with young children will quickly realize that the more novelty a toy displays the longer a child will remain interested in playing with it. If it is possible to create a toy which is able to change its behaviour over time or to adapt its behaviour to the user then it will enrich the child's experience with the toy and thereby enhance the learning process. To this end we examine important issues surrounding the embedding of CL algorithms in such a toy.