Coevolving functions in genetic programming: classification using K-nearest-neighbour

In this paper we introduce a new approach to the use of automatically defined functions within Genetic Programming for classification tasks. The technique consists of coevolving a number of separate sub-populations of functions, each of which acts as a feature extractor for the K-nearest-neighbour algorithm. Using two well-known classification tasks it is shown that our coevolutionary approach performs better than the equivalent traditional function mechanism used in Genetic Programming. The approach is then extended to include explicit feature selection at the level above the coevolving extractor functions. In this way features which are not needed for the task can be ignored more effectively than relying on the evolution of extractors which achieve the same effect. It is shown that this approach performs better than the first coevolutionary technique, and hence better than the traditional approach.