Leonardo Vanneschi

Genetic programming needs better benchmarks

Genetic programming (GP) is not a field noted for the rigor of its benchmarking. Some of its benchmark problems are popular purely through historical contingency, and they can be criticized as too easy or as providing misleading information concerning real-world performance, but they persist largely because of inertia and the lack of good alternatives. Even where the problems themselves are impeccable, comparisons between studies are made more difficult by the lack of standardization. We argue that the definition of standard benchmarks is an essential step in the maturation of the field. We make several contributions towards this goal. We motivate the development of a benchmark suite and define its goals; we survey existing practice; we enumerate many candidate benchmarks; we report progress on reference implementations; and we set out a concrete plan for gathering feedback from the GP community that would, if adopted, lead to a standard set of benchmarks.

An Empirical Study of Multipopulation Genetic Programming

This paper presents an experimental study of distributed multipopulation genetic programming. Using three well-known benchmark problems and one real-life problem, we discuss the role of the parameters that characterize the evolutionary process of standard panmictic and parallel genetic programming. We find that distributing individuals between subpopulations offers in all cases studied here an advantage both in terms of the quality of solutions and of the computational effort spent, when compared to single populations. We also study the influence of communication patterns such as the communication topology, the number of individuals exchanged and the frequency of exchange on the evolutionary process. We empirically show that the topology does not have a marked influence on the results for the test cases studied here, while the frequency and number of individuals exchanged are related and there exists a suitable range for those parameters which is consistently similar for all the problems studied.

A Study of Fitness Distance Correlation as a Difficulty Measure in Genetic Programming

We present an approach to genetic programming difficulty based on a statistical study of program fitness landscapes. The fitness distance correlation is used as an indicator of problem hardness and we empirically show that such a statistic is adequate in nearly all cases studied here. However, fitness distance correlation has some known problems and these are investigated by constructing an artificial landscape for which the correlation gives contradictory indications. Although our results confirm the usefulness of fitness distance correlation, we point out its shortcomings and give some hints for improvement in assessing problem hardness in genetic programming.

A survey of semantic methods in genetic programming

Several methods to incorporate semantic awareness in genetic programming have been proposed in the last few years. These methods cover fundamental parts of the evolutionary process: from the population initialization, through different ways of modifying or extending the existing genetic operators, to formal methods, until the definition of completely new genetic operators. The objectives are also distinct: from the maintenance of semantic diversity to the study of semantic locality; from the use of semantics for constructing solutions which obey certain constraints to the exploitation of the geometry of the semantic topological space aimed at defining easy-to-search fitness landscapes. All these approaches have shown, in different ways and amounts, that incorporating semantic awareness may help improving the power of genetic programming. This survey analyzes and discusses the state of the art in the field, organizing the existing methods into different categories. It restricts itself to studies where semantics is intended as the set of output values of a program on the training data, a definition that is common to a rather large set of recent contributions. It does not discuss methods for incorporating semantic information into grammar-based genetic programming or approaches based on formal methods. The objective is keeping the community updated on this interesting research track, hoping to motivate new and stimulating contributions.

A new implementation of geometric semantic GP and its application to problems in pharmacokinetics

Moraglio et al. have recently introduced new genetic operators for genetic programming, called geometric semantic operators. These operators induce a unimodal fitness landscape for all the problems consisting in matching input data with known target outputs (like regression and classification). This feature facilitates genetic programming evolvability, which makes these operators extremely promising. Nevertheless, Moraglio et al. leave open problems, the most important one being the fact that these operators, by construction, always produce offspring that are larger than their parents, causing an exponential growth in the size of the individuals, which actually renders them useless in practice. In this paper we overcome this limitation by presenting a new efficient implementation of the geometric semantic operators. This allows us, for the first time, to use them on complex real-life applications, like the two problems in pharmacokinetics that we address here. Our results confirm the excellent evolvability of geometric semantic operators, demonstrated by the good results obtained on training data. Furthermore, we have also achieved a surprisingly good generalization ability, a fact that can be explained considering some properties of geometric semantic operators, which makes them even more appealing than before.

Measuring bloat, overfitting and functional complexity in genetic programming

Recent contributions clearly show that eliminating bloat in a genetic programming system does not necessarily eliminate overfitting and vice-versa. This fact seems to contradict a common agreement of many researchers known as the minimum description length principle, which states that the best model is the one that minimizes the amount of information needed to encode it. Another common agreement is that overfitting should be, in some sense, related to the functional complexity of the model. The goal of this paper is to define three measures to respectively quantify bloat, overfitting and functional complexity of solutions and show their suitability on a set of test problems including a simple bidimensional symbolic regression test function and two real-life multidimensional regression problems. The experimental results are encouraging and should pave the way to further investigation. Advantages and drawbacks of the proposed measures are discussed, and ways to improve them are suggested. In the future, these measures should be useful to study and better understand the relationship between bloat, overfitting and functional complexity of solutions.

Fitness Clouds and Problem Hardness in Genetic Programming

This paper presents an investigation of genetic programming fitness landscapes. We propose a new indicator of problem hardness for tree-based ge- netic programming, called negative slope coefficient, based on the concept of fitness cloud. The negative slope coefficient is a predictive measure, i.e. it can be calculated without prior knowledge of the global optima. The fitness cloud is gener- ated via a sampling of individuals obtained with the Metropolis-Hastings method. The reliability of the negative slope coefficient is tested on a set of well known and representative genetic programming benchmarks, comprising the binomial-3 problem, the even parity problem and the artificial ant on the Santa Fe trail.

Genetic programming for computational pharmacokinetics in drug discovery and development

The success of a drug treatment is strongly correlated with the ability of a molecule to reach its target in the patient’s organism without inducing toxic effects. Moreover the reduction of cost and time associated with drug discovery and development is becoming a crucial requirement for pharmaceutical industry. Therefore computational methods allowing reliable predictions of newly synthesized compounds properties are of outmost relevance. In this paper we discuss the role of genetic programming in predictive pharmacokinetics, considering the estimation of adsorption, distribution, metabolism, excretion and toxicity processes (ADMET) that a drug undergoes into the patient’s organism. We compare genetic programming with other well known machine learning techniques according to their ability to predict oral bioavailability (%F), median oral lethal dose (LD50) and plasma-protein binding levels (%PPB). Since these parameters respectively characterize the percentage of initial drug dose that effectively reaches the systemic blood circulation, the harmful effects and the distribution into the organism of a drug, they are essential for the selection of potentially good molecules. Our results suggest that genetic programming is a valuable technique for predicting pharmacokinetics parameters, both from the point of view of the accuracy and of the generalization ability.

Negative slope coefficient: a measure to characterize genetic programming fitness landscapes

Negative slope coefficient has been recently introduced and empirically proven a suitable hardness indicator for some well known genetic programming benchmarks, such as the even parity problem, the binomial-3 and the artificial ant on the Santa Fe trail. Nevertheless, the original definition of this measure contains several limitations. This paper points out some of those limitations, presents a new and more relevant definition of the negative slope coefficient and empirically shows the suitability of this new definition as a hardness measure for some genetic programming benchmarks, including the multiplexer, the intertwined spirals problem and the royal trees.

Fitness distance correlation in structural mutation genetic programming

A new kind of mutation for genetic programming based on the structural distance operators for trees is presented in this paper. We firstly describe a new genetic programming process based on these operators (we call it structural mutation genetic programming). Then we use structural distance to calculate the fitness distance correlation coefficient and we show that this coefficient is a reasonable measure to express problem difficulty for structural mutation genetic programming for the considered set of problems, i.e. unimodal trap functions, royal trees and MAX problem.