A methodology for processing problem constraints in genetic programming

Search mechanisms of artificial intelligence combine two elements: representation, which determines the search space, and a search mechanism, which actually explores the space. Unfortunately, many searches may explore redundant and/or invalid solutions. Genetic programming refers to a class of evolutionary algorithms based on genetic algorithms, but utilizing a parameterized representation in the form of trees. These algorithms perform searches based on simulation of nature. They face the same problems of redundant/invalid subspaces. These problems have just recently been addressed in a systematic manner. This paper presents a methodology devised for the public domain genetic programming tool lil-gp. This methodology uses data typing and semantic information to constrain the representation space so that only valid, and possibly unique, solutions will be explored. The user enters problem-specific constraints, which are transformed into a normal set. This set is checked for feasibility, and subsequently, it is used to limit the space being explored. The constraints can determine valid, possibly unique spaces. Moreover, they can also be used to exclude subspaces the user considers uninteresting, using some problem-specific knowledge. A simple example is followed thoroughly to illustrate the constraint language, transformations, and the normal set. Experiments with Boolean 11-multiplexer illustrate practical applications of the method to limit redundant space exploration by utilizing problem-specific knowledge.