Enhancing Letter Recognition with Adaptive Classifiers
Exploring machine rule induction for a challenging letter recognition task using a Holland-style classifier system. A diverse set of 20,000 unique letter images was created by distorting pixel images of uppercase letters from various fonts. Different font styles were represented, including script, italic, serif, and Gothic. Each letter image was described by 16 primitive numerical attributes. Our focus was on machine induction techniques to create IF-THEN classifiers based on attribute values, with the THEN part indicating the correct letter category. We investigated the impact of attribute encoding methods, rule generation procedures, and credit allocation strategies. Binary and Gray-code encodings requiring exact matches for rule activation were evaluated.
文件大小：1.36MB
评论区