Raúl Pérez

Completeness and consistency conditions for learning fuzzy rules

The completeness and consistency conditions were introduced in order to achieve acceptable concept recognition rules. In real problems, we can handle noise-affected examples and it is not always possible to maintain both conditions. Moreover, when we use fuzzy information there is a partial matching between examples and rules, therefore the consistency condition becomes a matter of degree. In this paper, a learning algorithm based on soft consistency and completeness conditions is proposed. This learning algorithm combines in a single process rule and feature selection and it is tested on different databases.

Selection of relevant features in a fuzzy genetic learning algorithm

Genetic algorithms offer a powerful search method for a variety of learning tasks, and there are different approaches in which they have been applied to learning processes. Structural learning algorithm on vague environment (SLAVE) is a genetic learning algorithm that uses the iterative approach to learn fuzzy rules. SLAVE can select the relevant features of the domain, but when working with large databases the search space is too large and the running time can sometimes be excessive. We propose to improve SLAVE by including a feature selection model in which the genetic algorithm works with individuals (representing individual rules) composed of two structures: one structure representing the relevance status of the involved variables in the rule, the other one representing the assignments variable/value. For this general representation, we study two alternatives depending on the information coded in the first structure. When compared with the initial algorithm, this new approach of SLAVE reduces the number of rules, simplifies the structure of the rules and improves the total accuracy.

Including a simplicity criterion in the selection of the best rule in a genetic fuzzy learning algorithm

Learning algorithms can obtain very useful descriptions of several problems. Many different alternative descriptions can be generated. In many cases, a simple description is preferable since it has a higher possibility of being valid in unseen cases and also it is usually easier to understand by a human expert. Thus, the main idea of this paper is to propose simplicity criteria and to include them in a learning algorithm. In this case, the learning algorithm will reward the simplest descriptions. We study simplicity criteria in the selection of fuzzy rules in the genetic fuzzy learning algorithm called SLAVE.

Fuzzy Control of HVAC Systems Optimized by Genetic Algorithms

This paper presents the use of genetic algorithms to develop smartly tuned fuzzy logic controllers dedicated to the control of heating, ventilating and air conditioning systems concerning energy performance and indoor comfort requirements. This problem has some specific restrictions that make it very particular and complex because of the large time requirements existing due to the need of considering multiple criteria (which enlarges the solution search space) and to the long computation time models require to assess the accuracy of each individual.To solve these restrictions, a genetic tuning strategy considering an efficient multicriteria approach has been proposed. Several fuzzy logic controllers have been produced and tested in laboratory experiments in order to check the adequacy of such control and tuning technique. To do so, accurate models of the controlled buildings (two real test sites) have been provided by experts. Finally, simulations and real experiments were compared determining the effectiveness of the proposed strategy.

Three new instance selection methods based on local sets

The local set is the largest hypersphere centered on an instance such that it does not contain instances from any other class. Due to its geometrical nature, this structure can be very helpful for distance-based classification, such as classification based on the nearest neighbor rule. This paper is focused on instance selection for nearest neighbor classification which, in short, aims to reduce the number of instances in the training set without affecting the classification accuracy. Three instance selection methods based on local sets, which follow different and complementary strategies, are proposed. In an experimental study involving 26 known databases, they are compared with 11 of the most successful state-of-the-art methods in standard and noisy environments. To evaluate their performances, two complementary approaches are applied, the Pareto dominance relation and the Technique for Order Preference by Similarity to Ideal Solution. The results achieved by the proposals reveal that they are among the most effective methods in this field. HighlightsWe propose three selection strategies with different accuracy-reduction tradeoff.We assess them on 26 known databases with more than 1000 instances each one.The results are compared with those of 11 successful state-of-the-art methods.According to different criteria, the new methods are always among the top performers.

A study about the inclusion of linguistic hedges in a fuzzy rule learning algorithm

A very important problem associated with the use of learning algorithms consists of fixing the correct assignment of the initial domains for the predictive variables. In the fuzzy case, this problem is equivalent of define the fuzzy labels for each variable. In this work, we propose the inclusion in a learning algorithm, called SLAVE, of a particular kind of linguistic hedges as a way to modify the intial semantic of the labels. These linguistic hedges allow us both to learn and to tune fuzzy rules.

Combining instance selection methods based on data characterization: An approach to increase their effectiveness

Although there are several proposals in the instance selection field, none of them consistently outperforms the others over a wide range of domains. In recent years many authors have come to the conclusion that data must be characterized in order to apply the most suitable selection criterion in each case. In light of this hypothesis, herein we propose a set of measures to characterize databases. These measures were used in decision rules which, given their values for a database, select from some pre-selected methods, the method, or combination of methods, that is expected to produce the best results. The rules were extracted based on an empirical analysis of the behaviors of several methods on several data sets, then integrated into an algorithm which was experimentally evaluated over 20 databases and with six different learning paradigms. The results were compared with those of five well-known state-of-the-art methods.

A Set of Complexity Measures Designed for Applying Meta-Learning to Instance Selection

In recent years, some authors have approached the instance selection problem from a meta-learning perspective. In their work, they try to find relationships between the performance of some methods from this field and the values of some data-complexity measures, with the aim of determining the best performing method given a data set, using only the values of the measures computed on this data. Nevertheless, most of the data-complexity measures existing in the literature were not conceived for this purpose and the feasibility of their use in this field is yet to be determined. In this paper, we revise the definition of some measures that we presented in a previous work, that were designed for meta-learning based instance selection. Also, we assess them in an experimental study involving three sets of measures, 59 databases, 16 instance selection methods, two classifiers, and eight regression learners used as meta-learners. The results suggest that our measures are more efficient and effective than those traditionally used by researchers that have addressed the instance selection from a perspective based on meta-learning.

Knowledge-based instance selection: a compromise between efficiency and versatility

Traditionally, each instance selection proposal applies the same selection criterion to any problem. However, the performance of such criteria depends on the input data and a single one is not sufficient to guarantee success over a wide range of environments. An option to adapt the selection criteria to the input data is the use of meta-learning to build knowledge-based systems capable to choose the most appropriate selection strategy among several available candidates. Nevertheless, there is not in the literature a theoretical framework that guides the design of instance selection techniques based on meta-learning. This paper presents a framework for this purpose as well as a case study in which the framework is instantiated and an experimental study is carried out to show that the meta-learning approach offers a good compromise between efficiency and versatility in instance selection.

Overview of the SLAVE learning algorithm: A review of its evolution and prospects

AbstractInductive learning has been—and still is—one of the most important methods that can be applied in classification problems. Knowledge is usually represented using rules that establish relationships between the problem variables. SLAVE (Structural Learning Algorithm in a Vague Environment) was one of the first fuzzy-rule learning algorithms, and since its first implementation in 1994 it has been frequently used to benchmark new algorithms. Over time, the algorithm has undergone several modifications, and identifying the different versions developed is not an easy task. In this work we present a study of the evolution of the SLAVE algorithm from 1996 to date, marking the most important landmarks as definitive versions. In order to add these final versions to the KEEL platform, Java implementations have been developed. Finally, we describe the parameters used and the results obtained in the experimental study.