A stepwise evolutionary approach to machine learning

The biological concepts of stepwise evolution and cumulative selection are used to develop programs for a digital computer. The initial goal of these programs is to separate small, two-dimensional arrays of binary valued patterns into two classes. Learning is accomplished by the automatic evaluation of a criterion of separability of the two classes for each of the possible single-step mutations of the initial classification program. The mutant program which performs best is then selected automatically by the environment program and used as the new classification program. Only four operations are initially available, but as each classification program is mutated and cumulatively selected based on its performance, a new classification program is developed for implementing this particular classification. Each resulting program is then assigned a next sequential number, stored in mass storage, and made available for use as a single program step by subsequent programs. As the development program is run, it not only learns how to dichotomize more patterns but also has available to it the results of previous evolutionary learning experiences.<>