Parsimony pressure, the explicit penalization of larger programs, has been increasingly used as a means of controlling code growth in genetic programming. However, in many cases parsimony pressure degrades the performance of the genetic program. In this paper we show that poor average results with parsimony pressure are a result of "failed" populations that overshadow the results of populations that incorporate parsimony pressure successfully. Additionally, we show that the effect of parsimony pressure can be measured by calculating the relationship between program size and performance within the population. This measure can be used as a partial indicator of success or failure for individual populations. Keywords Code growth, code bloat, parsimony, genetic programming, introns. 1. Introduction The use of parsimony pressure as a means of controlling the size of programs generated with genetic programming (GP) has grown considerably in recent years. In many cases parsimony pr...

This work it is intended to discover classification rules for diagnosing certain pathologies. These rules are capable of discriminating among 12 different pathologies, whose main symptom is chest pain. In order to discover these rules it was used genetic programming as well as some concepts of data mining, particularly the emphasis on the discovery of comprehensible knowledge. 1 INTRODUCTION In order to classify and diagnose some pathology, one must verify which predicting attributes are most associated with that disease. In this work there are 189 predicting attributes and 12 different diseases (classes) whose main characteristic is chest pain. The predicting attributes refer to characteristics of the chest pain, other symptons reported by the patient, signals observed by the physician, details of clinical history and results of laboratory tests. The diseases are: stable angina, unstable angina, acute myocardial infarction, aortic dissection, cardiac tamponade, pulmonary emb...

Theorem proving essentially amounts to solving search problems. The intricacy of these in general undecidable problems makes the use of appropriate search-guiding heuristics indispensable. However, the appropriateness of a heuristic critically depends on the problem to be solved. Given a set of heuristics to choose from, selecting a suitable heuristic is hence a crucial, but also a very difficult task. It is usually taken care of by a proficient user, because it is very hard to determine the suitability of a certain heuristic based on a given problem to be solved. We propose here to automate the selection of heuristics using machine-learning techniques which ground their decisions on past problem-solving experience. Experimental studies conducted in a very difficult area of theorem proving, namely equational reasoning, demonstrate the capacity of the techniques and underline their potential to be a very useful tool for eliminating human interaction requiring expert knowledge....

In this paper, a novel feature extraction method is proposed; Genetic Programming (GP) is used to discover features, while the Fisher criterion is employed to provide fitness values. This produces nonlinear features for both two-class and multi-class recognition problems by revealing the discriminating information between classes. The proposed approach is experimentally compared to conventional nonlinear feature extraction methods, including kernel generalised discriminant analysis (KGDA), kernel principal component analysis (KPCA). Results demonstrate the capability of the proposed approach to transform information from the high dimensional feature space into a single dimensional space by automatically discovering the relationships among data. 1.

Abstract. We present a multi-dimensional mapping strategy using multiobjective genetic programming (MOGP) to search for the (near-)optimal feature extraction pre-processing stages for pattern classification as well as optimizing the dimensionality of the decision space. We search for the set of mappings with optimal dimensionality to project the input space into a decision space with maximized class separability. The steady-state Pareto converging genetic programming (PCGP) has been used to implement this multi-dimensional MOGP. We examine the proposed method using eight benchmark datasets from the UCI database and the Statlog project to make quantitative comparison with conventional classifiers. We conclude that MMOGP outperforms the comparator classifiers due to its optimized feature extraction process. 1

Automated reasoning or theorem proving essentially amounts to solving search problems. Despite significant progress in recent years theorem provers still have many shortcomings. The use of machine-learning techniques is acknowledged as promising, but difficult to apply in the area of theorem proving. We propose here to learn search-guiding heuristics by employing features in a simple, yet effective manner. Features are used to adapt a heuristic to a solved source problem. The adapted heuristic can then be utilized profitably for solving related target problems. Experiments have demonstrated that the approach allows a theorem prover to prove hard problems that were out of reach before. This work was supported by the Deutsche Forschungsgemeinschaft (DFG). 1 2 1 INTRODUCTION 1 Introduction Automated deduction or theorem proving essentially amounts to solving search problems. In general these problems are undecidable. Therefore, the search spaces encountered in general are infi...

a case study in a medical domain

This Thesis addresses the task of feature construction for classification. The quality of the data is one of the most important factors influencing the performance of any classification algorithm. The attributes defining the feature space of a given data set can often be inadequate, making it difficult to discover interesting knowledge. However, even when the original attributes are individually inadequate, it is often possible to combine such attributes in order to construct new ones with greater predictive power. The goal of this Thesis is to restructure the feature space in order to improve the performance of decision tree classification techniques on complex, real world data. The proposed framework involves the use of genetic programming to evolve (construct) new attributes, which are non--linear combinations of the original attributes. This approach incorporates a number of decision tree splitting mechanisms in the fitness measures of the genetic program. The empirical

Data mining deals with the problem of discovering novel and interesting knowledge from large amount of data. This problem is often performed heuristically when the extraction of patterns is difficult using standard query mechanisms or classical statistical methods. In this paper a genetic programming framework, capable of performing an automatic discovery of classification rules easily comprehensible by humans, is presented. A comparison with the results achieved by other techniques on a classical benchmark set is carried out. Furthermore, some of the obtained rules are shown and the most discriminating variables are evidenced. © 2002 Elsevier Science B.V. All rights reserved.