Papers by Ester Bernadó Mansilla
2020 IEEE Global Engineering Education Conference (EDUCON)
This contribution presents a course on imagination held at Politecnico di Milano. The aim of the ... more This contribution presents a course on imagination held at Politecnico di Milano. The aim of the course was to make students reflect on the role of imagination in various contexts (especially scientific and moral) and to teach them how to strengthen it to better perform as whole engineers. The pilot initiative was highly valued by the students as an innovative and meaningful way to introduce the humanities into technical mindsets.
International Journal of Hybrid Intelligent Systems, 2009
XCS is a learning classifier system that uses genetic algorithms to evolve a population of classi... more XCS is a learning classifier system that uses genetic algorithms to evolve a population of classifiers online. When applied to classification problems described by continuous attributes, XCS has demonstrated to be able to evolve classification models-represented as a set of independent interval-based rules-that are, at least, as accurate as those created by some of the most competitive machine learning techniques such as C4.5. Despite these successful results, analyses of how the different genetic operators affect the rule evolution for the interval-based rule representation are lacking. This paper focuses on this issue and conducts a systematic experimental analysis of the effect of the different genetic operators. The observations and conclusions drawn from the analysis are used as a tool for designing new operators that enable the system to extract models that are more accurate than those obtained by the original XCS scheme. More specifically, the system is provided with a new discovery component based on evolution strategies, and a new crossover operator is designed for both the original discovery component and the new one based on evolution strategies. In all these cases, the behavior of the new operators are carefully analyzed and compared with the ones provided by original XCS. The overall analysis enables us to supply important insights into the behavior of different operators and to improve the learning of interval-based rules in real-world domains on average.
This paper presents Fuzzy-UCS, a Michigan-style Learning Fuzzy-Classifier System designed for sup... more This paper presents Fuzzy-UCS, a Michigan-style Learning Fuzzy-Classifier System designed for supervised learning tasks. Fuzzy-UCS combines the generalization capabilities of UCS with the good interpretability of fuzzy rules to evolve highly accurate and understandable rule sets. Fuzzy-UCS is tested on a set of real-world problems, and compared to UCS and two of the most used machine learning techniques: C4.5 and SMO. The results show that Fuzzy-UCS is highly competitive to the three learners in terms of performance, and that the fuzzy representation permits a much better understandability of the evolved knowledge. These promising results allow for further investigation on Fuzzy-UCS.
Proceedings of the 9th annual conference on Genetic and evolutionary computation - GECCO '07, 2007
In this paper, we derive models of the selection pressure in XCS for proportionate (roulette whee... more In this paper, we derive models of the selection pressure in XCS for proportionate (roulette wheel) selection and tournament selection. We show that these models can explain the empirical results that have been previously presented in the literature. We validate the models on simple problems showing that, (i) when the model assumptions hold, the theory perfectly matches the empirical evidence; (ii) when the model assumptions do not hold, the theory can still provide qualitative explanations of the experimental results.
Proceedings of the 9th annual conference on Genetic and evolutionary computation, 2007
This paper analyzes the scalability of the population size required in XCS to maintain niches tha... more This paper analyzes the scalability of the population size required in XCS to maintain niches that are infrequently activated. Facetwise models have been developed to predict the effect of the imbalance ratio-ratio between the number of instances of the majority class and the minority class that are sampled to XCS-on population initialization, and on the creation and deletion of classifiers of the minority class. While theoretical models show that, ideally, XCS scales linearly with the imbalance ratio, XCS with standard configuration scales exponentially. The causes that are potentially responsible for this deviation from the ideal scalability are also investigated. Specifically, the inheritance procedure of classifiers' parameters, mutation, and subsumption are analyzed, and improvements in XCS's mechanisms are proposed to effectively and efficiently handle imbalanced problems. Once the recommendations are incorporated to XCS, empirical results show that the population size in XCS indeed scales linearly with the imbalance ratio.
Lecture Notes in Computer Science, 2010
The landscape contest provides a new and configurable framework to evaluate the robustness of sup... more The landscape contest provides a new and configurable framework to evaluate the robustness of supervised classification techniques and detect their limitations. By means of an evolutionary multiobjective optimization approach, artificial data sets are generated to cover reachable regions in different dimensions of data complexity space. Systematic comparison of a diverse set of classifiers highlights their merits as a function of data complexity. Detailed analysis of their comparative behavior in different regions of the space gives guidance to potential improvements of their performance. In this paper we describe the process of data generation and discuss performances of several well-known classifiers as well as the contestants’ classifiers over the obtained data sets.
Lecture Notes in Computer Science, 2008
This paper presents a learning methodology based on a substructural classification model to solve... more This paper presents a learning methodology based on a substructural classification model to solve decomposable classification problems. The proposed method consists of three important components: (1) a structural model, which represents salient interactions between attributes for a given data, (2) a surrogate model, which provides a functional approximation of the output as a function of attributes, and (3) a classification model, which predicts the class for new inputs. The structural model is used to infer the functional form of the surrogate. Its coefficients are estimated using linear regression methods. The classification model uses a maximally-accurate, least-complex surrogate to predict the output for given inputs. The structural model that yields an optimal classification model is searched using an iterative greedy search heuristic. Results show that the proposed method successfully detects the interacting variables in hierarchical problems, groups them in linkages groups, and builds maximally accurate classification models. The initial results on non-trivial hierarchical test problems indicate that the proposed method holds promise and also shed light on several improvements to enhance the capabilities of the proposed method.
Studies in Computational Intelligence, 2008
This chapter gives insight in the use of Genetic-Based Machine Learning (GBML) for supervised tas... more This chapter gives insight in the use of Genetic-Based Machine Learning (GBML) for supervised tasks. Five GBML systems which represent different learning methodologies and knowledge representations in the GBML paradigm are selected for the analysis: UCS, GAssist, SLAVE, Fuzzy AdaBoost, and Fuzzy LogitBoost. UCS and GAssist are based on a non-fuzzy representation, while SLAVE, Fuzzy Ad-aBoost, and Fuzzy LogitBoost use a linguistic fuzzy representation. The models evolved by these five systems are compared in terms of performance and interpretability to the models created by six highly-used nonevolutionary learners. Experimental observations highlight the suitability of GBML systems for classification tasks. Moreover, the analysis points out which systems should be used depending on whether the user prefers to maximize the accuracy or the interpretability of the models. fuzzy representation is that it allows for a better interpretability of the knowledge evolved, providing a flexible, robust, and powerful methodology to deal with noisy, imprecise, and incomplete data. The aim of this work is to evaluate the performance and the interpretability of the models evolved by five different GBML architectures in data mining tasks, and to compare their behavior to other non-evolutionary techniques. We include representatives of the different tendencies of GBML for supervised learning in the comparison. We select two GBML systems that use a non-fuzzy (crisp) representation: UCS [6], and GAssist [4]; and three GBML methods based on a fuzzy representation: SLAVE [29], Fuzzy AdaBoost [14], and Fuzzy LogitBoost . We compare these GBML systems to six highly-used non-evolutionary techniques. These learners come from different learning paradigms such as instance-based learning, rule and decision-tree induction, statistical modeling, and neural networks. The algorithms are compared on a collection of twenty real-world datasets extracted from the UCI repository [8] and local repositories . The remaining of this paper is organized as follows. Section 2 presents the different approaches in GBML, and Sect. 3 briefly explains the five GBML systems selected for the comparison. Section 4 details the experimentation comparison and present the results. Finally, Sect. 5 summarizes and concludes the work.
Proceedings of the 10th annual conference companion on Genetic and evolutionary computation, 2008
This paper presents CSar, a Michigan-style Learning Classifier System which has been designed for... more This paper presents CSar, a Michigan-style Learning Classifier System which has been designed for extracting quantitative association rules from streams of unlabeled examples. The main novelty of CSar with respect to the existing association rule miners is that it evolves the knowledge on-line and so it is prepared to adapt its knowledge to changes in the variable associations hidden in the stream of unlabeled data quickly and efficiently. Preliminary results provided in this paper show that CSar is able to evolve interesting rules on problems that consist of both categorical and continuous attributes. Moreover, the comparison of CSar with Apriori on a problem that consists only of categorical attributes highlights the competitiveness of CSar with respect to more specific learners that perform enumeration to return all possible association rules. These promising results encourage us for further investigating on CSar.
Studies in Computational Intelligence, 2008
Studies in Computational Intelligence, 2008
This chapter investigates the capabilities of XCS for mining imbalanced datasets. Initial experim... more This chapter investigates the capabilities of XCS for mining imbalanced datasets. Initial experiments show that, for moderate and high class imbalances, XCS tends to evolve a large proportion of overgeneral classifiers. Theoretical analyses are developed, deriving an imbalance bound up to which XCS should be able to differentiate between accurate and overgeneral classifiers. Some relevant parameters that have to be properly configured to satisfy the bound for high class imbalances are detected. Configuration guidelines are provided, and an algorithm that automatically tunes these XCS's parameters is presented. Finally, XCS is tested on a large set of real-world problems, appearing to be highly competitive to some of the most well-known machine learning techniques.
Lecture Notes in Computer Science, 2009
One of the most important challenges in supervised learning is how to evaluate the quality of the... more One of the most important challenges in supervised learning is how to evaluate the quality of the models evolved by different machine learning techniques. Up to now, we have relied on measures obtained by running the methods on a wide test bed composed of real-world problems. Nevertheless, the unknown inherent characteristics of these problems and the bias of learners may lead to inconclusive results. This paper discusses the need to work under a controlled scenario and bets on artificial data set generation. A list of ingredients and some ideas about how to guide such generation are provided, and promising results of an evolutionary multi-objective approach which incorporates the use of data complexity estimates are presented.
Lecture Notes in Computer Science
This paper introduces an approximate fuzzy representation to Fuzzy-UCS, a Michigan-style Learning... more This paper introduces an approximate fuzzy representation to Fuzzy-UCS, a Michigan-style Learning Fuzzy-Classifier System that evolves linguistic fuzzy rules, and studies whether the flexibility provided by the approximate representation results in a significant improvement of the accuracy of the models evolved by the system. We test Fuzzy-UCS with both approximate and linguistic representation on a large collection of real-life problems and compare the results in terms of training and test accuracy and interpretability of the evolved rule sets.
Lecture Notes in Computer Science, 2008
This paper presents Fuzzy-UCS, a Michigan-style Learning Fuzzy-Classifier System designed for sup... more This paper presents Fuzzy-UCS, a Michigan-style Learning Fuzzy-Classifier System designed for supervised learning tasks. Fuzzy-UCS combines the generalization capabilities of UCS with the good interpretability of fuzzy rules to evolve highly accurate and understandable rule sets. Fuzzy-UCS is tested on a large collection of real-world problems, and compared to UCS and three highly-used machine learning techniques: the decision tree C4.5, the support vector machine SMO, and the fuzzy boosting algorithm Fuzzy LogitBoost. The results show that Fuzzy-UCS is highly competitive with respect to the four learners in terms of performance, and that the fuzzy representation permits a much better understandability of the evolved knowledge. These promising results of the online architecture of Fuzzy-UCS allow for further research and application of the system to new challenging problems.
Proceedings of the 12th annual conference on Genetic and evolutionary computation, 2010
Currently available real-world problems do not cover the whole complexity space and, therefore, d... more Currently available real-world problems do not cover the whole complexity space and, therefore, do not allow us to thoroughly test learner behavior on the border of its domain of competence. Thus, the necessity of developing a more suitable testing scenario arises. With this in mind, data complexity analysis has shown promise in characterizing difficulty of classification problems through a set
2008 Eighth International Conference on Hybrid Intelligent Systems, 2008
In this paper, we highlight the use of synthetic data sets to analyze learners behavior under bou... more In this paper, we highlight the use of synthetic data sets to analyze learners behavior under bounded complexity. We propose a method to generate synthetic data sets with a specific complexity, based on the length of the class boundary. We design a genetic algorithm as a search technique and find it useful to obtain class labels according to the desired complexity. The results show the suitability of the genetic algorithm as a framework to provide artificial benchmark problems that can be further enriched with the use of multi-objective and niching strategies.
Lecture Notes in Computer Science, 2001
ABSTRACT MOLeCS is a classifier system (CS) which addresses its learning as a multiobjective task... more ABSTRACT MOLeCS is a classifier system (CS) which addresses its learning as a multiobjective task. Its aim is to develop an optimal set of rules, optimizing the accuracy and the generality of each rule simultaneously. This is achieved by considering these two goals in the rule fitness. The paper studies four multiobjective strategies that establish a compromise between accuracy and generality in different ways. The results suggest that including the decision maker’s preferences in the search process improves the overall performance of the obtained rule set. The paper also studies a third major objective: covering (the maintenance of a set of different rules solving together the learning problem), through different niching mechanisms. After a performance analysis using some benchmark problems, MOLeCS is applied to a real-world categorization task: the diagnosis of breast cancer.
Proceedings of the 8th annual conference on Genetic and evolutionary computation, 2006
This paper analyzes the behavior of the XCS classifier system on imbalanced datasets. We show tha... more This paper analyzes the behavior of the XCS classifier system on imbalanced datasets. We show that XCS with standard parameter settings is quite robust to considerable class imbalances. For high class imbalances, XCS suffers from biases toward the majority class. We analyze XCS's behavior under such extreme imbalances and prove that appropriate parameter tuning improves significantly XCS's performance. Specifically, we counterbalance the imbalance ratio by equalizing the reproduction probabilities of the most occurring and least occurring niches. The study provides guidelines to tune XCS's parameters for unbalanced datasets, based on the dataset imbalance ratio. We propose a method to estimate the imbalance ratio during XCS's training and adapt XCS's parameters online.
2005 IEEE Congress on Evolutionary Computation
The class imbalance problem has been said to challenge the performance of concept learning system... more The class imbalance problem has been said to challenge the performance of concept learning systems. Learning systems tend to be biased towards the majority class, and thus have poor generalization for the mi- nority class instances. We analyze the class imbalance problem in learning classifier systems based on genetic algorithms. In particular we study UCS, a rule-based classifier system which learns under a supervised learn- ing scheme. We analyze UCS on an artificial domain with varying imbalance levels. We find UCS fairly sen- sitive to high levels of class imbalance, to the degree that UCS tends to evolve a simple model of the feature space classified according to the majority class. We analyze strategies for dealing with class imbalances, and find fit- ness adaptation based on class-sensitive accuracy a use- ful tool for alleviating the effects of class imbalances. In the last decades, research in genetic algorithms (GAs) and evolutionary computation (EC) has paid increasing at- tention to machine learning and data mining applications. Particularly, classification has been one of the primary in- terests of researchers working in data mining applications of genetic algorithms. Classification can be defined as the process of assigning a class label to a given example, given a set of examples previously classified. Many approaches exist, such as decision trees, instance-based learners, neural networks, and others. Evolutionary learning classifier systems (LCSs) have demonstrated to be highly competitive with respect to other classifier schemes in a varied range of domains. Since the first proposal, developed by Holland [Hol75, Hol76], the field has benefited from numerous research and development, being XCS [Wil95, Wil98] one of the best representatives. At the current stage of maturity, researchers have started to analyze the domain of competence of LCSs [BH05], and tested LCSs on challenging real-world classi- fication problems [BerO2, BLGO2, BB04, ButO4]. Research on real-world domains has identified several sources of complexity for classifier schemes, such as the geometry of class boundaries, sparsity of the available train- ing dataset, presence of noise, and class imbalances, among others [KKOI, BH05]. Class imbalances correspond to the case where one class is represented by a larger number of in- stances than other classes. The issue is of great importance since it appears in many real-world domains, such as fraud
Abstract. XCS is a complex,machine,learning technique that combines,credit ap- portionment techni... more Abstract. XCS is a complex,machine,learning technique that combines,credit ap- portionment techniques for rule evaluation with genetic algorithms for rule discov- ery to evolve a distributed set of sub-solutions online. Recent research on XCS has mainly focused on achieving a better understanding of the reinforcement compo- nent, yielding several improvements to the architecture. Nonetheless, studies on the rule discovery component of the system are scarce. In this paper, we experimentally study the discovery component of XCS, which is guided by a steady-state genetic algorithm. We design a new procedure based on evolution strategies and adapt it to the system. Then, we compare in detail XCS with both genetic algorithms and evolution strategies on a large collection of real-life problems, analyzing in detail the interaction of the different genetic operators and their contribution in the search for better rules. The overall analysis shows the competitiveness of the new XCS based on ...
Uploads
Papers by Ester Bernadó Mansilla