PERFORMANCE-DRIVEN AUTONOMOUS DESIGN OF PATTERN-RECOGNITION SYSTEMS

The closed-loop design experiment described in this paper demonstrates a three-phase automated design approach to pattern recognition. The experiment generates morphological feature detectors and then uses a novel application of genetic algorithms to select cooperative sets of features to pass to a neural net classifier. The self-organizing hybrid learning approach embodied in this closed-loop design methodology is complementary to conventional artificial intelligence (AI) expert systems that utilize rule-based approaches and a specific set of design elements. This experiment is part of a study directed to emulating the nondirected processes of biological evolution. The approach we discuss is semiautomatic in that initialization of computer programs requires human experience and expertise to select representations, develop search strategies, choose performance measures, and devise resource-allocation strategies. The hope is that these tasks will become easier with experience and will provide the means to exploit parallel processing without the need to analyze or program an entire design solution.