Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity.However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning.Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification.We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient.We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts.Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool language models win up to 41.5% more often in pairwise comparisons of reasoning processes).This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes.Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity).Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tool outputs as assistive evidence, improving both final-answer accuracy and the reasoning depth under tool use. 1

Pouya Pezeshkpour, Estevam Hruschka. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Bosung Kim, Hayate Iso, Nikita Bhutani, Estevam Hruschka, Ndapa Nakashole, Tom Mitchell. Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

The exponentially growing of generated data and the advanced computing techniques, catalyzed the necessity of innovative strategies to store, analyze and capture biological data. The biocuration was created due to the number and scope of scientific databases in recent years. Biocuration is an essential part of biological and biomedical research discovery. Extracted from the literature, curated data is so important to accomplish computational analysis and to train data mining algorithms. Concurrently, the researchers need a rapidly and more decisive approach to understand unknown domains. In this paper, we propose an architecture to help biocurators in Gene Ontology (GO) classification tasks. Our approach is based on a semi-supervised environment that couples evolutionary computation (i.e. Genetic Algorithms (GA)) and traditional classifiers (i.e. Decision Trees and Naive Bayes). Furthermore, our approach provides high level knowledge (IF-THEN rules and decision trees), balancing accuracy, interpretability and comprehensibility that can help GO biocurators in their classification tasks. Our architecture shows higher classification rates (about 94%, i.e., 17,820 correct samples of 18,959), starting from a small training set, only 2,707 samples (12.5% of 21,666 samples).

User-Generated text is a rich source of user insights and experiences that can be very helpful in many different daily life situations, such as when deciding what product to buy, what hotel to stay, what company to apply for a job, what region to buy a house, etc. This kind of text also plays a very relevant role in current research efforts in academic research groups, technology companies, as well as big publishers, telecommunications players, recruiting and job-market focuses organizations, etc. The goal of this new workshop is to bring together researchers interested in the application of novel techniques in AI/ML/NLP and Knowledge Discovery to address challenges around harnessing text-heavy user-generated data that is available to organizations and over the Web. The workshop program contains invited speakers, contributed talks, poster sessions and a discussion panel

This tutorial focuses on how to build Question Answering (QA) syetems based on the Never-Ending Learning (NEL) approach. NEL systems can be roughly described as computer systems that learn over time to become better in solving a specific task. Different NEL approaches have been proposed and applied in different tasks and domains. Recent advances encourage us to keep addressing the problem of how to build computer systems that can take advantage of NEL principles. Considering that it is not always so straightforward to have NEL principles applied to ML models, this tutorial guides the audience (with hands-on examples and supporting theory, algorithms and models) on how to model a system in a NEL fashion and intends to help KDD community to become familiar with such approaches. Question Answering is chosen as application domain mainly because of the relevance of the topic (QA) for KDD and AI communities in general.

This tutorial introduces Lifelong Machine Learning (LML) and Machine Reading. The core idea of LML is to learn continuously and accumulate the learned knowledge, and to use the knowledge to help future learning, which is perhaps the hallmark of human learning and human intelligence. By us- ing prior knowledge seamlessly and effortlessly, we humans can learn without a lot of training data, but current machine learning algorithms tend to need a huge amount of training data. LML aims to mimic this human capability. Machine Reading is a research area with the goal of building systems to read natural language text. Among different approaches employed in Machine Reading, this tutorial focuses on projects and approaches that use the idea of LML. Most current machine learning (ML) algorithms learn in isolation. They are designed to address a specific problem using a single dataset. That is, given a dataset, an ML algorithm is executed on the dataset to build a model. Although this type of isolated learning is very useful, it does not have the ability to accumulate past knowledge and to make use of the knowledge for future learning, which we believe are critical for the future of machine learning and data mining. LML aims to design and develop computational systems and algorithms with this capability, i.e., to learn as humans do in a lifelong manner.

Supervised algorithms require a set of representative labeled data for building classification models. However, labeled data are usually difficult and expensive to obtain, which motivates the interest in semi-supervised learning. This type of learning uses both labeled and unlabeled data in the training process and is particularly useful in applications such as tweet sentiment analysis, where a large amount of unlabeled data is available. Semi-supervised learning for tweet sentiment analysis, although quite appealing, is relatively new. We propose a semi-supervised learning framework that combines unsupervised information, captured from a similarity matrix constructed from unlabeled data, with a classifier. Our motivation is that such a similarity matrix is a powerful knowledge-discovery tool that can help classify unlabeled tweet sets. Our framework makes use of the well-known Self-training algorithm to induce a better tweet sentiment classifier. Experimental results in real-world datasets demonstrate that the proposed framework can improve the accuracy of tweet sentiment analysis.

Whereas people learn many different types of knowledge from diverse experiences over many years, most current machine learning systems acquire just a single function or data model from just a single data set. We propose a never-ending learning paradigm for machine learning, to better reflect the more ambitious and encompassing type of learning performed by humans. As a case study, we describe the Never-Ending Language Learner (NELL), which achieves some of the desired properties of a never-ending learner, and we discuss lessons learned. NELL has been learning to read the web 24 hours/day since January 2010, and so far has acquired a knowledge base with over 80 million confidence-weighted beliefs (e.g., servedWith(tea, biscuits)). NELL has also learned millions of features and parameters that enable it to read these beliefs from the web. Additionally, it has learned to reason over these beliefs to infer new beliefs, and is able to extend its ontology by synthesizing new relational predicates. NELL can be tracked online at http://rtw.ml.cmu.edu, and followed on Twitter at @CMUNELL.

NELL (Never Ending Language Learning system) is the first system to practice the Never-Ending Machine Learning paradigm techniques. It has an inactive component to continually extend its KB: OntExt. Its main idea is to identify and add to the KB new relations which are frequently asserted in huge text data. Co-occurrence matrices are used to structure the normalized values of co-occurrence between the contexts for each category pair to identify those context patterns. The clustering of each matrix is done with Weka K-means algorithm: from each cluster, a new possible relation. This work present newOntExt: a new approach with new features to turn the ontology extension task feasible to NELL. This approach has also an alternative task of naming new relations found by another NELL component: Prophet. The relations are classified as valid or invalid by humans; the precision is calculated for each experiment and the results are compared to those relative to OntExt. Initial results show that ontology extension with newOntExt can help Never-Ending Learning systems to expand its volume of beliefs and to keep learning with high precision by acting in auto-supervision and auto-reflection.

The first Never-Ending Learning system reported in the literature, which is called NELL (Never-Ending Language Learner), was designed to perform the task of autonomously building an knowledge base as a result of continuously reading the web. NELL is based on a learning paradigm in which, the learner, in an autonomous way, manages to constantly, incrementally and continuously evolve with time. But, most important than just keep evolving, in this paradigm acquired knowledge is used, in a dynamic way, to expand the scope and improve the performance of the learning task as a whole. Coreference resolution plays a key role in any system based on the Never-Ending Learning paradigm. In this paper two diferente views of correference resolution are applied to NELL's knowledge base and empirical evidence is obtained to show that combining morphological and semantic features in a hybrid model can be more effective than using only one of the feature views.

In recent years, many researches have been focusing their studies in large growing knowledge bases. Most techniques focus on building algorithms to help the Knowledge Base (KB) automatically (or semi-automatically) extends. In this article, we make use of a generalized association rule mining algorithm in order, specially, to increase the relations between KB's categories. Although, association rules algorithms generates many rules and evaluate each one is a hard step. So, we also developed a structure, based on pruning obvious itemsets and generalized rules, which decreases the amount of discovered rules. The use of generalized association rules contributes to their reduction. Experiments confirm that our approach helps to increase the relationships between the KB's domains as well as facilitate the process of evaluating extracted rules.

An alternative to the traditional single function approximation method is the never-ending learning (NEL) approach i.e., a learning paradigm in which, the learner, in an autonomous way, manages to constantly, incrementally and continuously evolve with time. But, most important than just keep evolving, in this new paradigm acquired knowledge can, in a dynamic way, be used to expand the scope and improve the performance of the learning task as a whole. The first Never-Ending Learning system reported in the literature, which is called NELL (Never-Ending Language Learner), is applied to the task of autonomously building a knowledge base as a result of reading the web. Results reported so far reveal that very good results have been achieved when NELL is reading the web in English. When trying, however, to perform the same Machine Reading task (the task of reading the web) applied to web pages written in Portuguese, the previous reported approaches could not keep up with the good performance achieved in English. In this paper we describe an approach, different from previously proposed in the literature, and we present empirical results that corroborate the hypothesis that working on the preprocessing task of a sufficiently big corpus can be key to allow us to use the very same architecture proposed in NELL, but applied to the idea of reading the web in Portuguese (reading, and extracting knowledge from web pages written in Portuguese).

In this paper we present an approach to treatment of the Cold-Start Problem in Recommendation System for Environment Education Web. Our approach is based on the concept of Coupled-Learning and Bootstrapping. Based on an initial set of data we apply algorithms traditional machine learning to cooperate with each other, forming various views on its outputs and allowing the data set to be classified incrementally. Thus, it is possible to increase the initial volume of data and to improve the performance of a recommender more instances for analysis. The vast majority of the efforts attack the cold start problem with variations of the CBF algorithm. In our approach, we use the incremental semi-supervised learning based on pairs in order to increase the initial training set in order to allow the generation of more recommendations.

The project and implementation of autonomous computational systems that incrementally learn and use what has been learnt to, continually, refine its learning abilities throughout time is still a goal far from being achieved. Such dynamic systems woul

Time plays an important role in the vast majority of problems and, as such, it is a vital issue to be considered when developing computer systems for solving problems. In the literature, one of the most influential formalisms for representing time is

A number of different computational approaches have been applied in many different biology application domains. When such tools are based on conventional computation techniques, they have shown limitations to approach complex biological problems. In the present study, a genetic algorithm (named GANEL) that is based on some Never-Ending Learning (NEL) principles, is proposed as a tool to extract classification rules from biological datasets. The main goal of the proposed approach is to allow the discovery of concise, yet accurate, high-level rules (from a biological database) which can be used to describe the stronger patterns present in the biological data, revealing concise and relevant information about the application domain, as well as, be used as a classification system. More than focusing only on the classification accuracy, the proposed GANEL approach aims at balancing prediction precision, interpretability and comprehensibility. The obtained results show that the proposed GANEL is promising and capable of extracting useful high-level knowledge that could not be extracted by traditional classifications methods such as Decision Trees, One R and the Single Conjunctive Rule Learner, among others. Moreover, the accuracy of GANEL results (using a small set of attributes per class) are better than Computational Evolutionary Environment (CEE) (previously proposed in the literature) which was designed to the same problem domain.

This paper describes a proposal which extends Allen's interval algebra by adapting the formalism for dealing with binary relations involving time periods with uncertain boundaries. The extended formalism has been proposed having in mind the subsequent investigation of the automatic learning of temporal relations using an inductive logic programming system (FOIL). Learning results are presented and discussed.

In the 1950s and the 1960s several computer scientists independently studied evolutionary systems with the idea that evolution could be used as an optimization tool for engineering problems. For these evolutionary-computation researchers, the mechanisms of evolution seem well suited for some of the most pressing computational problems in many fields. Ideas from Genetics are usually incorporated into evolutionary algorithms, such as: haploid crossover, mutation, diploid, inversion, gene doubling, deletion, and others. In the present study, we proposed an operator, named transgenic, for evolutionary algorithms, especially designed for Genetic Algorithms (GA). This operator is inspired by genetically modified organisms (GMOs), where important features are introduced into their genome artificially. The transgenic operator uses historical information to choose the best attributes, converging to better results faster than traditional GAs. The GA, used in this study, allows the discovery of concise, yet accurate, high-level rules (from a biological and synthetic database) which can be used as a classification system. The obtained results show that transgenic operator is promising at obtaining better of the the same results with a lower number of generations and smaller populations.

Variable Ordering (VO) plays an important role when inducing Bayesian Networks (BNs). Previous works in the literature suggest that it is worth pursuing the use of evolutionary strategies for identifying a suitable VO, when learning a Bayesian Network structure from data. This paper proposes a hybrid adaptive algorithm named VOMOS (Variable Ordering Multiple Offspring Sampling) where the new individuals are created using a set of recombination operators (crossover and mutation operators). Experiments performed in datasets revealed that the VOMOS approach is promising and tends to generate consistent and representative BNs.

This work proposes and discusses an approach for inducing Bayesian classifiers aimed at balancing the tradeoff between the precise probability estimates produced by time consuming unrestricted Bayesian networks and the computational efficiency of Nai

We consider here the problem of building a never-ending language learner; that is, an intelligent computer agent that runs forever and that each day must (1) extract, or read, information from the web to populate a growing structured knowledge base, and (2) learn to perform this task better than on the previous day. In particular, we propose an approach and a set of design principles for such an agent, describe a partial implementation of such a system that has already learned to extract a knowledge base containing over 242,000 beliefs with an estimated precision of 74% after running for 67 days, and discuss lessons learned from this preliminary attempt to build a never-ending learning agent.

Computational approaches have been applied in many different biology application domains. When such tools are based on conventional computation, they have shown limitations to approach complex biological problems. In the present study, a computational evolutionary environment (CEE) is proposed as tool to extract classification rules from biological datasets. The main goal of the proposed approach is to allow the discovery of concise, yet accurate, high-level rules (from a biological database) which can be used as a classification system. More than focusing only on the classification accuracy, the proposed CEE model aims at balancing prediction precision, interpretability and comprehensibility. The obtained results show that the proposed CEE is promising and capable of extracting useful high-level knowledge that could not be extracted by traditional classifications methods such as Decision Trees, One R and the Single Conjunctive Rule Learner using the same dataset.

Variable Orderings (VOs) have been used as a restriction in the process of Bayesian Networks (BNs) induction. The VO information can significantly reduce the search space and allow some algorithms to reach good results. Previous works reported in the literature suggest that the combination of Evolutionary Algorithms (EAs) and VOs is worth when learning a Bayesian Network structure from data. However, most works on this area do not explore specific characteristics of the domain, thus, they simply apply classic evolutionary operators. In addition, most works did not report good results when applied to big BNs. This paper proposes a new mutation operator, named Distance-Based Mutation Operator (DMO), to be used with the Variable Ordering Evolutionary Algorithm (VOEA). Experimental results obtained by VOEA are compared to ones achieved by VOGA (Variable Ordering Genetic Algorithm), and indicated improvement in the quality of the obtained VO and in the BN induced structure.

Variable Ordering plays an important role when inducing Bayesian Networks. Previous works in the literature suggest that the use of genetic/evolutionary algorithms (EAs) for dealing with VO, when learning a Bayesian Network structure from data, is worth pursuing. This work proposes a new crossover operator, named Random Multi-point Crossover Operator (RMX), to be used with the Variable Ordering Evolutionary Algorithm (VOEA). Empirical results obtained by VOEA are compared to the ones achieved by VOGA (Variable Ordering Genetic Algorithm), and indicated improvement in the quality of VO and the induced BN structure.

Hybrid intelligent systems which take advantage of the Bayesian/Fuzzy collaboration have been explored in the literature in the last years. Such collaboration can play an important role mainly in real intelligent systems applications, where accuracy and comprehensibility are crucial aspects to be considered. This paper further explore the Bayes Fuzzy method proposing a classification method specially designed to be used in intelligent systems for data analysis. The main idea is to enhance comprehensibility while maintaining accuracy by decreasing the number of fuzzy rules used to explain a Bayesian Classifier (BC). The proposed Pruned Bayes Fuzzy 2 (PBF2) method is based on a new feature selection method named Selection by Markov Blanket Relation Strength (SMBRS). In the performed experiments, PBF2 is empirically applied to a real world police records problem in order to extract a comprehensible and accurate set of rules which can help in crime prevention. The obtained results show PBF2, when used with proper parameters, brings better precision and comprehensibility compared to other Bayesian/Fuzzy-based methods and to C4.5 algorithm.

We consider the problem of semi-supervised learning to extract categories (e.g., academic fields, athletes) and relations (e.g., PlaysSport(athlete, sport)) from web pages, starting with a handful of labeled training examples of each category or relation, plus hundreds of millions of unlabeled web documents. Semi-supervised training using only a few labeled examples is typically unreliable because the learning task is underconstrained. This paper pursues the thesis that much greater accuracy can be achieved by further constraining the learning task, by coupling the semi-supervised training of many extractors for different categories and relations. We characterize several ways in which the training of category and relation extractors can be coupled, and present experimental results demonstrating significantly improved accuracy as a result.

The use of Bayesian Network Classifiers (BCs) combined with the Fuzzy rule model to explain the learned BCs have been previously presented as the BayesFuzzy approach. This paper follows along BayesFuzzy lines of investigation aiming at improving the comprehensibility of a BC model and enhancing BayesFuzzy results by combining new pruning methods. In order to improve BayesFuzzy performance, in addition to the Markov Blanket-based pruning idea used by BayesFuzzy, two other pruning methods are proposed, implemented and empirically evaluated. The first pruning method is based on the conditional probability estimates given by the BC and the second one is the well-known post-rule pruning approach, usually used to prune rules extracted from decision trees. Also, three different Bayesian Networks induction algorithms, namely IC, K2 and Nai¿ve-Bayes, as well as, the C4.5 Decision Tree induction algorithms are employed in the empirical comparative analysis performed in the experiments. The obtained results reveal that BayesFuzzy combined with the new pruning methods can bring comprehensibility enhancements.

The substitution of missing values, also called imputation, is an important data preparation task for many domains. Ideally, the substitution of missing values should not insert biases into the dataset. This aspect has been usually assessed by some measures of the prediction capability of imputation methods. Such measures assume the simulation of missing entries for some attributes whose values are actually known. These artificially missing values are imputed and then compared with the original values. Although this evaluation is useful, it does not allow the influence of imputed values in the ultimate modelling task (e.g. in classification) to be inferred. We argue that imputation cannot be properly evaluated apart from the modelling task. Thus, alternative approaches are needed. This article elaborates on the influence of imputed values in classification. In particular, a practical procedure for estimating the inserted bias is described. As an additional contribution, we have used such a procedure to empirically illustrate the performance of three imputation methods (majority, naive Bayes and Bayesian networks) in three datasets. Three classifiers (decision tree, naive Bayes and nearest neighbours) have been used as modelling tools in our experiments. The achieved results illustrate a variety of situations that can take place in the data preparation practice.

This work proposes a genetic strategy for learning a Bayesian classifier using an algorithm based on conditional independence and the information given by a variable ordering. The strategy has been implemented as the system VOGAC-PC. The paper presents and analyses the results of experiments in various domains using VOGAC-PC as well as a previous system, named VOGA-K2, based on algorithm K2.

Traditional algorithms for learning Bayesian Classifiers (BCs) from data are known to induce accurate classification models. However, when using these algorithms, two main concerns should be considered: i) they require qualitative data and ii) generally the induced models are not easily comprehensible by human beings. This paper deals with the two above issues by proposing a hybrid method named BayesFuzzy that learns from quantitative data and induces a fuzzy rule based model that enhances comprehensibility. BayesFuzzy has been implemented as an automatic system that combines a fuzzy strategy, for transforming numerical data into qualitative information, with a Bayes-based approach for inducing rules. Promising empirical results of the use of the BayesFuzzy system in four knowledge domains are presented and discussed.

Bayesian network (BN) is a formalism for representing and reasoning about uncertain domains. In BN the knowledge is represented by a combination of a graph-based structure and probability theory. A particular type of BN known as Bayesian Classifier (BC) aims at classifying a given instance into a discrete class. BCs have been extensively used for modeling knowledge in many different applications and have been the focus of many works related to data mining. Depending on the size of a BC the understandability of the knowledge it represents is not an easy task. This paper proposes an approach to help the process of understanding the knowledge represented by a BC, by translating it into a more convenient and easily understandable form of representation, that of classification rules. The proposed method named BayesRule (BR) uses the concept of Markov Blanket (MB) to obtain a reduced set of rules in respect to both, the number of rules and the number of antecedents in rules. Experiments using the ALARM network showed that the reduced set of rules extracted from the BC can be smaller than the set of rules representing a decision tree generated by C4.5, and still maintains a high accuracy rate.

This work proposes, implements and discusses a hybrid Bayes/genetic collaboration (VOGAC-MarkovPC) designed to induce conditional independence Bayesian classifiers from data. The main contribution is the use of MarkovPC algorithm in order to reduce the computational complexity of a genetic algorithm (GA) designed to explore the variable orderings (VOs) in order to optimize the induced classifiers. Experiments performed in a number of datasets revealed that VOGAC-MarkovPC required less than 25% of the time demanded by VOGAC-PC on average. In addition, when concerning the classification accuracy, VOGAC-MakovPC performed as well as VOGAC-PC did.

This paper describes the modeling of a biomass based weed-crop competitiveness classification process, based on a Bayesian network classifier. The understandability of the model is improved by its automatic translation into a set of classification rules, which are easily understood by human beings. The Bayes approach is based on empirical data col- lected in a corn-crop and uses the concept of maximum a posteriori probability to extract a set of probabilistic rules from the induced Bayesian network classifier. The features used to build the Bayesian network classifier are the total density of weeds and the corresponding proportions of nar- row and broadleaf weeds and the class variable is the weeds biomass from which the weed-crop competitiveness is in- ferred. The paper presents a set of 27 rules extracted from the Bayesian network classifier which classify the biomass of weeds.

The definition of the fuzzy rule base is one of the most important and difficult tasks when designing fuzzy systems. This paper discusses the results of two different hybrid methods investigated earlier, for the automatic generation of fuzzy rules from numerical data. One of the methods proposes the creation of fuzzy rule bases using genetic algorithms in association with a heuristic for preselecting candidate rules. The other, named Bayes fuzzy, induces a Bayes classifier using a dataset previously granulated by fuzzy partitions and then translates this classifier into a fuzzy rule base. A comparative analysis between both approaches focusing on their main characteristics, strengths/weaknesses and easiness of use is carried out. The reliability of both methods is also compared by analyzing their results in a few knowledge domains.

The Naïve Bayes Classifier (NB) aims at classifying a given instance into a discrete class considering that all attributes are conditionally independent given the class. NB has been extensively used for modeling knowledge in many different applications and has been the focus of many works related to classification tasks. This work proposes and discusses a Naïve Bayesian classifier named Weighted Naïve Bayesian (WNB) classifier. The central idea of WNB is that more relevant attributes should have more influence in the classification estimation process. A weighting strategy is adopted to modify the traditional NB. Experiments performed with six UCI domains show that WNB is promising.

The problem of feature selection is particularly important in areas such as machine learning and data mining. Given a training set which generally is described as a set of instances, each of them represented as a vector of feature-value pairs and an associated class, a feature selection method tries to identify features that are relevant for describing the concept embedded in the training set. The Focus algorithm is a very popular feature subset approach used in discrete data; its version for continuous data is known as C-Focus. The selection process implemented by Focus searches for the smallest feature subset that does not produce inconsistency in the training set. Due to its search mechanism, however, the use of the Focus (or C-Focus) becomes restrictive when the number of features is very large. This paper proposes a heuristic search method to be used by Focus (C-Focus), aiming at improving its performance. Empirical results of a system implementing C-Focus with the new search strategy (C-Focus-3) show that the proposed heuristic performs very well when the relevant subset of features has more than half of the number of features present in the original training set.

Constructive neural network algorithms enable the architecture of a neural network to be constructed as an intrinsic part of the learning process. These algorithms are very dependent on the TLU training algorithm they employ. Generally they use a Perceptron-based algorithm (such as Pocket or Pocket with Ratchet Modification (PRM)) for training each individual node added to the network, during the learning process. In the literature can be found a vast selection of algorithms for training individual TLUs. This paper investigates the use of the Barycentric Correction Procedure (BCP) algorithm with four constructive algorithms namely Tower, Pyramid, Shift and Perceptron-Cascade. Results show that some constructive neural algorithms have better performance using BCP than using PRM.

There are two fundamentally different approaches for feature selection: wrapper and filter. It is also possible to combine them, obtaining hybrid approaches. This paper describes a hybrid method for selecting relevant features in clustering problems. The proposed approach is based on the combination of the widely known k-means algorithm and a Bayesian filter, which is based on the Markov Blanket concept. Since the number of clusters and the subset of relevant features are usually inter-related, we propose a method that iterates between clustering (assuming that the number of clusters is not known a priori) and filtering. Experiments in a number of datasets show that the proposed approach allows selecting features that provide good partitions.

This paper proposes and evaluates a hybrid system based on two machine learning approaches, a neural network and an instance based method. It describes how the knowledge induced by a RuleNet neural network can be used as the initial knowledge for an NGE-like system to start learning. An NGE-based system can be considered an instance based learning method which allows generalization. The proposed collaboration between the two learning methods implemented by the hybrid system is feasible due to the similarity of the concept description languages employed by both. The paper also describes a few experiments conducted; results show that the RuleNet-NGE collaboration is plausible and, in some domains, it improves the performance of NGE on its own.

As a previous step to machine learning (ML) induced classifiers, attribute subset selection methods have become an efficient alternative for reducing the dimensionality of the search space, with obvious benefits to the learning techniques used. This paper investigates the problem of feature subset selection using a committee of filter, wrapper and embedded methods. The wrappers were implemented using two different search mechanisms, a genetic algorithm and a best-first procedure as well as three different machine learning paradigms: instance-based (nearest neighbor - NN), neural network (DistAl) and symbolic (C4.5). The two filter methods used are based on consistency and correlation measures. The goals of the experiments were to be able to identify the most suitable attribute subsets to be further used for inducing a classifier as well as investigate if the combination of different results given by the committee's members can outperform any machine learning method using the original training set.