Applications of learning strategies to pattern recognition

摘要

Abstract: Experiments are described with a hybrid learning systemthat automates the generation of multiclass patternrecognition systems. The learning system utilizesgenetic algorithms to formulate a small set of featuredetectors and an adaptive neural network to classifyfeature vectors. The experiments utilize a training setof handprinted characters and a pool of randomlygenerated morphological hit-or-miss detectors. A majorproblem is selecting a small subset of cooperatingdetectors from a large, easily generated pool ofdetectors. A new method of selecting detectors from apool is presented that utilizes a modified version ofcrossover operators from genetic algorithms. This newcrossover approach for generating feature detectors isevaluated by comparing it with random selection andwith a restricted random mutation approach. The systemadaptively adjusts the size of detector sets and thecorresponding number of neural net nodes.!