Adding new features and classes to classifiers evolved using genetic programming

This paper considers the need to re-train a multiclass classifier that has initially been evolved using genetic programming to accommodate new features or new classes. For the former, the new feature is incorporated into a program by mutation; after that, a program that performs classification using all the features is obtained by evolution. For the latter, a binary classifier is evolved that is able to distinguish the new class from all existing classes is evolved, and it is executed before the existing classification programs. The two approaches are demonstrated on a range of classification problems drawn from the general area of produce grading and the results demonstrate the effectiveness of the proposed approach, in terms of both computational speed and classification performance.