Letter Recognition Using Holland-Style Adaptive Classifiers

摘要: Machine rule induction was examined on a difficult categorization problem by applying Holland-style classifier system to complex letter recognition task. A set of 20,000 unique images generated randomly distorting pixel the 26 uppercase letters from 20 different commercial fonts. The parent fonts represented full range character types including script, italic, serif, and Gothic. features each characters were summarized in terms 16 primitive numerical attributes. Our research focused machine techniques for generating IF-THEN classifiers which IF part list values attributes THEN correct category, i.e., one alphabet. We effects procedures encoding attributes, deriving new rules, apportioning credit among rules. Binary Gray-code attribute encodings that required exact matches activation compared with integer representations employed fuzzy matching activation. Random genetic methods creation instance-based generalization. strength/specificity method apportionment procedure we call “accuracy/utility.“