Genetic Programming(GP) is gaining increased attention as an effective method for inducing classifiers for data classification. However, the manual design of a genetic programming classification algorithm is a non-trivial time consuming process. This research investigates the hypothesis that automating the design of a GP classification algorithm for data classification can still lead to the induction of effective classifiers and also reduce the design time. Two evolutionary algorithms, namely, a genetic algorithm (GA) and grammatical evolution (GE) are used to automate the design of GP classification algorithms. The classification performance of the automated designed GP classifiers i.e. GA designed GP classifiers and GE designed GP classifiers are compared to each other and to manually designed GP classifiers on real-world problems. Furthermore, a comparison of the design times of automated design and manual design is also carried out for the same set of problems. The automated designed classifiers were found to outperform manually designed classifiers across problem domains. Automated design time is also found to be less than manual design time. This study revealed that for the considered datasets GE performs better for binary classification while the GA does better for multiclass classification. Overall the results of the study are in support of the hypothesis.

Introduction

Data classification, is one of the most widely studied domains of research in machine learning. Many real-world tasks can be viewed as classification problems. Classification is the process of associating an object to a class (label) based 5 on the features describing that object. Classification is generally performed by classifier models. Classification usually involves two phases a learning (training) phase and a testing phase. A classifier model is induced by a classification algorithm during training and its classification accuracy is evaluated during testing. Evolutionary algorithms(EAs) are one of the methods that have gained 10 prominence in the induction of classifiers, particularly genetic programming (Espejo et al. (2010)Freitas (2003)). Genetic programming(GP) is a population based algorithm that models Darwin’s theory of evolution(Koza (1992)). For a number of reasons GP has proved to be effective in the induction of classifiers. The tree representation used by GP allows it flexibility to evolve classifiers that 15 model numerous problems(Espejo et al. (2010)). For example GP can be configured to represent decision trees, association rules or discriminant functions. GP like most EAs is a parameterized algorithm and it has been shown that the effectiveness of such algorithms depends on their configuration (Eiben and Smit (2011)). Algorithm configuration is a design process that involves deter20 mining numerical parameter values, selecting categorical parameters and setting the control flow that would lead to the algorithm finding an optimal (or near optimal) solution to the problem at hand. According to Kramer and Kacprzyk (2008) to manually configure an EA that yields effective results, considerable algorithm design experience is necessary. However Hutter (2009) argues that 25 even with the necessary experience manual design is a tedious non-trivial task susceptible to human bias. Furthermore, the search space for possible parameter values is large and only a subset of the design decisions are considered during manual design. Parameter values and flow control options are considered using trial runs in an iterative trial and error approach. Montero and Riff 30 (2014) argue that inexperienced designers add more parameters than are necessary during manual design leading to unnecessarily complex algorithms. They also point out that a lot of man hours are required for manual design and ideally the algorithm designer should have expert knowledge of the domain being considered, however this is not always possible. Parameter control and tuning 35 methods have been proposed in literature (Dobslaw (2010); Eiben et al. (1999); Eiben and Smit (2011)) with no method being universally adopted (Karafotias et al. (2015)).