Catálogo de publicaciones - libros

In the paper, a new method of decision tree learning for cost-sensitive classification is presented. In contrast to the traditional greedy top-down inducer in the proposed approach optimal trees are searched in a global manner by using an evolutionary algorithm (EA). Specialized genetic operators are applied to modify both the tree structure and tests in non-terminal nodes. A suitably defined fitness function enables the algorithm to minimize the misclassification cost instead of the number of classification errors. The performance of the EA-based method is compared to three well-recognized algorithms on real-life problems with known and randomly generated cost-matrices. Obtained results show that the proposed approach is competitive both in terms of misclassification cost and compactness of the classifier at least for some datasets.

In a paper [1] presented to BICS 2006, a basic methodology for microprocessor design automation using DNA sequences was proposed. A refined methodology with new schemes for traversal, encoding, recombination, and processor evaluation are proposed in this paper. Moreover concepts such as mutation, graphical decoding and environment simulation are introduced and a new technique for creating DNA based algorithms used in the mutation process is also presented. The proposed methodology is then generalized to extend its application to other domains. This paper presents a conceptual framework whose implementation aspects are still under investigation.

A new evolutionary-progressive method for Multiple Sequence Alignment problem is proposed. The method efficiently combines flexibility of evolutionary approach with speed and accuracy of progressive technique. The results show that the hybrid method is an interesting alternative for purely genetic or purely progressive approaches.

In many situations a robust design could be expensive and decision-makers need to evaluate a design that is not robust, that is, a design with a probability of satisfying the design specifications (or yield) less than 100 %. In this paper we propose a procedure for centring a design that maximises the yield, given predefined component tolerances. The hybrid approach is based on the use of Evolutionary Algorithms, Interval Arithmetic and procedures to estimate the yield percentage. The effectiveness of the method is tested on a literature case. We compare the special evolutionary strategy (1+1) with a genetic algorithm and deterministic, statistical and interval-based procedures for yield estimation.

In this paper we propose a new self-adaptative crossover operator for real coded evolutionary algorithms. This operator has the capacity to simulate other real-coded crossover operators dynamically and, therefore, it has the capacity to achieve exploration and exploitation dynamically during the evolutionary process according to the best individuals. In other words, the proposed crossover operator may handle the generational diversity of the population in such a way that it may either generate additional population diversity from the current one, allowing exploration to take effect, or use the diversity previously generated to exploit the better solutions.

In order to test the performance of this crossover, we have used a set of test functions and have made a comparative study of the proposed crossover against other classic crossover operators. The analysis of the results allows us to affirm that the proposed operator has a very suitable behavior; although, it should be noted that it offers a better behavior applied to complex search spaces than simple ones.

This paper provides two methodologies for forecasting time series. One of them is based on the Wavelet Analysis and the other one on the Genetic Programming. Two examples from finance domain are used to illustrate how given methodologies perform in real-life applications. Additionally application to specific classes of time series, seasonal, is discussed.

Evolutionary algorithms have been adequately applied in solving single and multi-objective optimization problems. In the single-objective case various studies have shown the usefulness of combining gradient based classical search principles with evolutionary algorithms. However there seems to be a dearth of such studies for the multi-objective case. In this paper, we take two classical search operators and discuss their use as a local search operator in a state-of-the-art evolutionary algorithm. These operators require gradient information which is obtained using a stochastic perturbation technique requiring only two function evaluations. Computational studies on a number of test problems of varying complexity demonstrate the efficiency of hybrid algorithms in solving a large class of complex multi-objective optimization problems.

Co-evolutionary techniques for evolutionary algorithms allow for the application of such algorithms to problems for which it is difficult or even impossible to formulate explicit fitness function. These techniques also maintain population diversity, allows for speciation and help overcoming limited adaptive capabilities of evolutionary algorithms. In this paper the idea of is introduced. In presented system the Pareto frontier is located by the population of agents as a result of co-evolutionary interactions between two species: predators and prey. Results from runs of presented system against test problem and comparison to classical multi-objective evolutionary algorithms conclude the paper.

Significant improvement over a patented lens design is achieved using multi-objective evolutionary optimization. A comparison of the results obtained from NSGA2 and -MOEA is done. In our current study, -MOEA converged to essentially the same Pareto-optimal solutions as the one with NSGA2, but -MOEA proved to be better in providing reasonably good solutions, comparable to the patented design, with lower number of lens evaluations. -MOEA is shown to be computationally more efficient and practical than NSGA2 to obtain the required initial insight into the objective function trade-offs while optimizing large and complex optical systems.

Radial Basis Function Neural Networks (RBFNNs) have been widely used to solve classification and regression tasks providing satisfactory results. The main issue when working with RBFNNs is how to design them because this task requires the optimization of several parameters such as the number of RBFs, the position of their centers, and their radii. The problem of setting all the previous values presents many local minima so Evolutionary Algorithms (EAs) are a common solution because of their capability of finding global minima. Two of the most important elements in an EAs are the crossover and the mutation operators. This paper presents a comparison between a non distributed multiobjective algorithm against several parallel approaches that are obtained by the specialisation of the crossover and mutation operators in different islands. The results show how the creation of specialised islands that use different combinations of crossover and mutation operators could lead to a better performance of the algorithm by obtaining better solutions.