Scalability analysis of genetic programming classifiers

Genetic programming (GP) has been used extensively for classification due to its flexibility, interpretability and implicit feature manipulation. There are also disadvantages to the use of GP for classification, including computational cost, bloating and parameter determination. This work analyses how GP-based classifier learning scales with respect to the number of examples in the classification training data set as the number of examples grows, and with respect to the number of features in the classification training data set as the number of features grows. The scalability of GP with respect to the number of examples is studied analytically. The results show that GP scales very well (in linear or close to linear order) with the number of examples in the data set and the upper bound on testing error decreases. The scalability of GP with respect to the number of features is tested experimentally, with results showing that the computations increase exponentially with the number of features.