GP made faster with semantic surrogate modelling

Genetic Programming (GP) is known to be expensive in cases where the fitness evaluation is computationally demanding, i.e., object detection, programmatic compression, image processing applications. The paper introduces a method that reduces the amount of fitness evaluations that are required to obtain good solutions. We consider the supervised learning setting, where a training set of input vectors are collectively mapped to a vector of outputs, and then a loss function is used to map the vector of outputs to a scalar fitness value. Saving of fitness evaluations is achieved through the use of two components. The first component is surrogate model that predicts trees output for a particular input vector xi based on the similarity between xi and other input vectors in the training set for which the candidate solution has been already evaluated with. The second component, is a simple linear equation to control the size of a sub-training set that is used to train GP trees. This linear equation allows the size of the sub-training set to dynamically increase or decrease based on the status of the search. The proposed method referred to as SSGP. Empirical results in 17 different problems, from three different categories, demonstrate that SSGP is able to obtain solutions of similar quality with those obtained using several benchmark GP systems, but with a much smaller computation time. The simplicity of the proposed method and the ease of its implementation is one of the most appealing aspects of its future utility.