Jaroslaw Arabas

GAVaPS-a genetic algorithm with varying population size

The size of the population can be critical in many applications of genetic algorithms. If the population size is too small, the genetic algorithm may converge too quickly; if it is too large, the genetic algorithm may waste computational resources; the waiting time for an improvement might be too long. We propose an adaptive method for maintaining variable population size, which grows and shrinks together according to some characteristic of the search. The first experimental results indicate some merits of the proposed method.<<ETX>>

Genetic Algorithms for the 0/1 Knapsack Problem

In this paper the utility of several constraint-handling techniques is investigated on the basis of a family of 0/1 knapsack problems. Several evolutionary algorithms are applied to this NP-hard problem. The conclusions might be applicable to many constrained combinatorial optimization problems, for which the use of evolutionary algorithm is considered.

Applying an evolutionary algorithm to telecommunication network design

This paper deals with the application of evolutionary computation to telecommunication network design. Design of a two-layer network is considered, where the upper-layer (UL) network uses resources of the lower-layer (LL) network. UL links determine demands for the LL and are implemented using LL paths (admissible paths). Within a fixed LL network topology, given the demands and admissible paths, we aim to find the LL link capacities for implementing the UL links, minimizing the cost of the LL. Robust design issues are also taken into consideration, allowing for failure of a certain part of the LL and postulating that, after some re-allocation in the LL, demands are still realized to an assumed extent. An algorithm based on an evolutionary technique is introduced, with problem-specific genetic operators to improve computing efficiency. A theoretical study of properties of the operators is made and several experiments are performed to tune the parameters of the algorithm. Finally, its performance is compared with other design techniques, including integer programming.

Differential mutation based on population covariance matrix

In this paper we analyze the impact of mutation schemes using many difference vectors in Differential Evolution (DE) algorithm. We show that for an infinite (sufficiently large) number of difference vectors, distribution of their sum weakly converges to a normal distribution. This facilitates theoretical analysis of DE and leads to introduction of a mutation scheme generalizing differential mutation using multiple difference vectors. The novel scheme uses Gaussian mutation with covariance matrix proportional to the covariance matrix of the current population instead of calculating difference vectors directly. Such modification, called DE/rand/∞, and its hybridization with DE/best/1 were tested on the CEC 2005 benchmark and performed comparable or better than DE/rand/1. Both modified mutation schemes may be easily incorporated into other DE variants. In this paper we provide theoretical analysis, discussion of obtained mutation distributions, and experimental results.

Approximating the Genetic Diversity of Populations in the Quasi-Equilibrium State

This paper analyzes an evolutionary algorithm in the quasi-equilibrium state, i.e., when the population of chromosomes fluctuates around a single peak of the fitness function. The analysis is aimed at approximating the genetic variance of the population when chromosomes are real-valued. The infinite population model is considered which allows the quasi-equilibrium state to be defined as the state when the density of chromosomes contained by the population remains unchanged over consecutive generations. This paper provides formulas for genetic diversity in the quasi-equilibrium state for fitness proportionate, tournament, and truncation selection types, with and without elitism, with Gaussian mutation, and with and without arithmetic crossover. The formulas are experimentally validated.

DMEA — An algorithm that combines differential mutation with the fitness proportionate selection

In this paper we report an ongoing work on an algorithm called DMEA (Evolutionary Algorithm based on the Differential Mutation). This algorithm is composed of differential mutation coupled with the “traditional” Gaussian mutation, fitness proportionate selection and generational replacement. We experimentally show that the DMEA is capable to generate chromosomes in a way that their distribution fits to the contour lines of the fitness function. Performance of DMEA was evaluated on the CEC2005 benchmark. Quality of results is comparable to many leading global optimization methods, including those which are based on the Differential Evolution paradigm.

A differential evolution strategy

This contribution introduces an evolutionary algorithm (EA) for continuous optimization in ℝn. The algorithm generates new individuals by the standard nonelitist truncation selection and the differential mutation to generate new individuals. The differential mutation is enriched by adding a random vector in the direction of the shift of population midpoint. Difference vectors are generated with the use of the archive of previous populations. Boundary constraints are handled by penalty function.

KIS: An automated attribute induction method for classification of DNA sequences

Abstract This paper presents an application of methods from the machine learning domain to solving the task of DNA sequence recognition. We present an algorithm that learns to recognize groups of DNA sequences sharing common features such as sequence functionality. We demonstrate application of the algorithm to find splice sites, i.e., to properly detect donor and acceptor sequences. We compare the results with those of reference methods that have been designed and tuned to detect splice sites. We also show how to use the algorithm to find a human readable model of the IRE (Iron-Responsive Element) and to find IRE sequences. The method, although universal, yields results which are of quality comparable to those obtained by reference methods. In contrast to reference methods, this approach uses models that operate on sequence patterns, which facilitates interpretation of the results by humans.

Benchmarking IBHM method using NN3 competition dataset

We apply a novel black box approximation algorithm, called IBHM, to learn both structure and parameters of a nonlinear regression model. IBHM incrementally creates a model as a weighted sum of activation functions which are nonlinear functions of the input vector. In each iteration the error between the current model and the approximated function is analyzed and a function is selected with the highest possible correlation with the observed error. This function is then added to the set of the models activation functions and the process repeats. In effect IBHM determines both the model structure and parameter values. In this paper we briefly outline the method and present the results on the NN3 benchmark set. We compare results with other state-of-the-art methods that share a similar model structure: Multilayer Perceptron with a single hidden layer and Support Vector Regression.

Differential Evolution: A survey of theoretical analyses

Abstract Differential Evolution (DE) is a state-of-the art global optimization technique. Considerable research effort has been made to improve this algorithm and apply it to a variety of practical problems. Nevertheless, analytical studies concerning DE are rather rare. This paper surveys the theoretical results obtained so far for DE. A discussion of genetic operators characteristic of DE is coupled with an overview of the population diversity and dynamics models. A comprehensive view on the current-day understanding of the underlying mechanisms of DE is complemented by a list of promising research directions.