Predicción de rendimiento y dificultad de problemas en programación genetica

Abstract

The estimation of problem difficulty is an open issue in Genetic Programming(GP). The goal of this work is to generate models that predictthe expected performance of a GP-based classifier when it is applied toan unseen task. Classification problems are described using domainspecificfeatures, some of which are proposed in this work, and thesefeatures are given as input to the predictive models. These models arereferred to as predictors of expected performance (PEPs). We extendthis approach by using an ensemble of specialized predictors (SPEP),dividing classification problems into groups and choosing the correspondingSPEP. The proposed predictors are trained using 2D syntheticclassification problems with balanced datasets. The models are thenused to predict the performance of the GP classifier on unseen realworlddatasets that are multidimensional and imbalanced. This workis the first to provide a performance prediction of a GP system on testdata, while previous works focused on predicting training performance.Accurate predictive models are generated by posing a symbolic regressiontask and solving it with GP. These results are achieved by usinghighly descriptive features and including a dimensionality reductionstage that simplifies the learning and testing process. The proposed approachcould be extended to other classification algorithms and usedas the basis of an expert system for algorithm selection.