Some experiments in machine learning using vector evaluated genetic algorithms (artificial intelligence, optimization, adaptation, pattern recognition)

This dissertation describes experiments conducted to explore the efficacy of using vector-valued feedback with a class of adaptive procedures called genetic algorithms. The software system developed was called VEGA for Vector Evaluated Genetic Algorithm and was first used on multiple objective optimization problems. The principle conclusion of these experiments was that VEGA provided a powerful and robust search technique for complex multiobjective optimization problems of high order when little or no a priori knowledge was available to guide the search. These results were similar to those found by previous researchers using scalar genetic algorithms for scalar optimization problems. The VEGA technique was then applied to multiclass pattern discrimination tasks. The resulting software system was called LS-2 for Learning System - Two since it followed closely the lead of a scalar-valued learning system called LS-1 developed by Stephen Smith. The experiments revealed that LS-2 was able to evolve high performance production system programs to perform the pattern discrimination tasks it was given. In addition, experiments which varied several of the parameters of LS-2 revealed something of the sensitivity of vector-valued genetic search to the settings of these parameters. In sum it may be said that the VEGA approach has demonstrated the efficacy of extending the previously demonstrated power of genetic algorithms to vector-valued problems and thereby provides a new approach to machine learning.