Processing Symbolic Data With Self-Organizing Maps

Current research on machine learning and related algorithms is focused mainly on numeric paradigms. It is, however, widely supposed that intelligent behavior strongly relies on ability to manipulate symbols. In this paper we present on-line versions of self-organizing maps for symbol strings. The underlying key concepts are average and similarity, applied on strings. These concepts are easily defined for numerical data and form building blocks for many learning, adaptation and clustering algorithms. By defining them for symbol strings, we propose to apply current algorithms to symbolic data.