Cooperative Co-Evolutionary Genetic Programming for High Dimensional Problems

We propose a framework for Cooperative Co-Evolutionary Genetic Programming (CCGP) that considers co-evolution at three different abstraction levels: genotype, feature and output level. A thorough empirical evaluation is carried out on a real-world high dimensional ML problem (image denoising). Results indicate that GP’s performance is enhanced only when cooperation happens at an output level (ensemble-alike). The proposed co-evolutionary ensemble approach is compared against a canonical GP implementation and a GP customized for image processing tasks. Preliminary results show that the proposed framework obtains superior average performance in comparison to the other GP models. Our most relevant finding is the empirical evidence showing that the proposed CCGP model is a promising alternative to specialized GP implementations that require knowledge of the problem’s domain.