Studying elements ofgenetic programming for multiclass classification

Tese de mestrado, Engenharia Informática (Interação e Conhecimento) Universidade de Lisboa, Faculdade de Ciências, 2018Although Genetic Programming (GP) has been very successful in both symbolic regression and binary classification by solving many difficult problems from various domains, it requires improvements in multiclass classification, which due to the high complexity of this kind of problems, requires specialized classifiers. In this project, we explored a multiclass classification GP-based algorithm, the M3GP [4]. The individuals in standard GP only have one node at their root. This means that their output space is in R. Unlike standard GP, M3GP allows each individual to have n nodes at its root. This variation changes the output space to Rn, allowing them to construct clusters of samples and use a cluster-based classification. Although M3GP is capable of creating interpretable models while having competitive results with state-of-the-art classifiers, such as Random Forests and Neural Networks, it has downsides. The focus of this project is to improve the algorithm by exploring two components, the fitness function, and the genetic operators’ selection method. The original fitness function was accuracy-based. Since using this kind of functions does not allow a smooth evolution of the output space, we tried to improve the algorithm by exploring two distance-based fitness functions as an attempt to separate the clusters while bringing the samples closer to their respective centroids. Until now, the genetic operators in M3GP were selected with a fixed probability. Since some operators have a better effect on the fitness at different stages of the evolution, the fixed probabilities allow operators to be selected at the wrong stages of the evolution, slowing down the learning process. In this project, we try to evolve the probability the genetic operators have of being chosen over the generations. On a later stage, we proposed a new crossover genetic operator that uses three individuals for the M3GP algorithm. The results obtained show significantly better results in the training set in half the datasets, while improving the test accuracy in two datasets