Genetic engineering of handwriting representations

This paper presents experiments with genetically engineered feature sets for recognition of on-line handwritten characters. These representations stem from a nondescript decomposition of the character frame into a set of rectangular regions, possibly overlapping, each represented by a vector of 7 fuzzy variables. Efficient new feature sets are automatically discovered using genetic programming techniques. Recognition experiments conducted on isolated digits of the Unipen database yield improvements of more than 3% over a previously manually designed representation where region positions and sizes were fixed.