A Memetic Genetic Programming with decision tree-based local search for classification problems

Abstract

Abstract-In this work, we propose a new genetic programming algorithm with local search strategies, named Memetic Genetic Programming(MGP), for classification problems. MGP aims to acquire a classifier with large Area Under the ROC Curve (AUC), which has been proved to be a better performance metric for traditionally used metrics (e.g., classification accuracy). Three new points are presented in our new algorithm. First, a new representation called statistical genetic decision tree (SGDT) for GP is proposed on the basis of Genetic Decision Tree (GDT). Second, a new fitness function is designed by using statistic information from SGDT. Third, the concept of memetic computing is introduced into SGDT. As a result, the MGP is equipped with a local search method based on the training algorithms for decision trees. The efficacy of the MGP is empirically justified against a number of relevant approaches.

Frequency Fitness Assignment

“…We integrated FFA into the Memetic Genetic Programming (MGP) system developed by Wang et al [27]. Evolving classifiers with GP exhibits less epistasis, ruggedness, and neutrality than the synthesis of deterministic algorithms with discrete, non-approximative results but still is a computational hard problem.…”

“…The MGP system additionally has another feature that makes it interesting as an application area for FFA: it is a highly specialized, fine-tuned tool with perfectly matched components. We will now outline the original MGP system briefly (while pointing to [27] for more details).…”

“…MGP [27] is used to synthesize tree-based classifiers. Usually, a decision tree contains a set of decision nodes that are traversed starting from its root for each data sample to be classified until a leaf node is reached that contains a class assignment.…”

“…Usually, a decision tree contains a set of decision nodes that are traversed starting from its root for each data sample to be classified until a leaf node is reached that contains a class assignment. MGP, however, builds statistical genetic decision trees (SGDTs) where the leaf nodes l instead hold an estimated probability p(l, k) for each class k [27] as illustrated in Figure 6 and Equation 3. These probabilities are learned during the training phase.…”

“…In the test/application phase, these frequencies determine which class the leave node assigns to a sample arriving at it. More information is given in [27].…”