The standard paradigm for exploring hypothesis testing behaviour is Wason's (1960) rule discovery task, which exists in two variants: the standard single goal (SG) task, and the logically identical dual goal (DG) fonn. Despite the close similarity of the two fonns of the task, the reported success rates in the two variants vary considerably, with approximately 20% of participants successfully solving the SG variant compared to over 60% correctly announcing the rule in the DG fonn. It was this disparity between the patterns of perfonnance across the two versions of the task which fonned the impetus for this thesis, as it was felt that an explanation for the facilitatory effect of DG instructions would lead to insights into the poor performance in the SG form. Several competing contemporary accounts of the effect are introduced, and predictions derived from them empirically tested across a series of seven experiments. Data analyses showed that no single contemporary theory could provide a wholly adequate account of the DG facilitation effect. However, these analyses led to a novel observation: that it is the production of a contrast class triple which appears to be the key predictor of success on the task, and furthennore, that the DG variant of the task promotes the generation of such a triple. Support for the "contrast class" account of the DG effect was provided by direct manipulation of the information provided to participants. A theoretical account of the critical role of contrast class cue information is developed in the thesis by situating the account within a proposed extension to Oaksford and Chater's (1994) "Iterative Counterfactual Model" of hypothesis testing. It is further suggested that rather than providing mutually exclusive accounts of the DG effect, competing theories (e.g., Vallee-Tourangeau et al. 's, 1995, triple heterogeneity theory, and Wharton et al. 's, 1993, information quantity theory) could be subsumed within this new model, which would then reflect a process whereby participants' strategies change and develop over the course of the hypothesis testing session. Finally, it is suggested that findings from this thesis can be accommodated more generally within Evans' (2006) "hypothetical thinking framework", and thereby within contemporary dual process accounts of reasoning.

L'agrégation d'activités pour l'identification de catégories de comportements est un enjeu majeur de tous les systèmes socio-techniques complexes actuels. La question clé consiste à réaliser une synthèse de façons de faire (ou praxies) intégrant la variabilité des opérateurs humains impliqués. Dans un cadre aéronautique, l'agrégation d'activités de pilotage vise à accélérer la détermination de procédures améliorant la sécurité des vols et l'efficacité des missions ; elle repose sur les données objectives des paramètres enregistrés des phases de vol significatives et se structure grâce à une interprétation experte. Un modèle d’Agrégation Supervisée : - décomposition, - maïeutique, - reconstruction, est ainsi établi dans cette thèse. Le cœur en est la 2e étape qui généralise et enrichit le concept de « moyenne » classique des approches probabilistes : une base d'apprentissage, constituée d'activités déterminées et caractérisées par l'interprétation experte, est utilisée pour identifier les motifs significatifs de paramètres enregistrés, c'est à dire les praxies qui agrègent donc les éléments essentiels des activités. Ceux-ci sont choisis au sein d'un ensemble de motifs paramétrables génériques, dont les divers seuils sont ajustés de manière incrémentale. Les motifs sont alors évalués selon les deux critères intrinsèques de cohérence et de pertinence de leurs seuils, ainsi que le critère extrinsèque de la conformité des résultats obtenus par leur utilisation aux vols de la base d'apprentissage. Peuvent à ce niveau se faire jour des groupements parmi les éléments de la base d'apprentissage, selon les motifs rendant compte des activités particulières. L'expertise doit également être généralisable pour permettre l'étude de plusieurs points-clé dans cette étape maïeutique.Ce modèle générique définit une activité comme une structure formelle de praxies, et ouvre la voie à un enrichissement de la 3e étape intégrant la multiplicité des rôles des opérateurs. / Aggregating activities in order to identify categories of behaviour is a major topic of actual complex socio-technical systems. The key issue lies in incorporating the variability of implied human operators in the synthesis of ways of doing (or praxis). Aggregation of piloting activities is directed to allow a faster and more secure determination of procedures enhancing flight security and mission efficiency; it is based on the objective data of flight parameters recorded during significant flight phases, and is carried under thorough expert interpretation.A Supervised Aggregation model, consisting in the 3 steps of 1) decomposition, 2) maieutics, and 3) reconstruction, is thus devised in the present PhD. At the heart of this aggregation process, the 2nd maieutic step generalizes and enriches the usual concept of ''mean'', deeply related to probabilistic approaches: a set of activities analyzed and characterized by the expert, the learning basis, is related to significant patterns in the lot of recorded flight parameter values, in other words the praxis resulting of the aggregation of the activities. The patterns are selected from a collection of customizable generic patterns, whose thresholds are incrementally adjusted using the learning basis. The obtained patterns are then assessed according to the three criteria of 1) coherence and 2) likelihood of the thresholds, as well as the 3) conformity of these patterns used on the learning basis. At this stage, groups among the studied behaviours might emerge, gathering those for which an activity would be depicted by similar patterns. Expert-knowledge must be generalized in order to perform the joint analysis of several key points in this maieutic step.This generic model defines an activity as a formal structure of praxis, paving the way towards the further developments of the process, through the enrichment of the 3rd step, incorporating the multiplicity of operating roles.

Na área de inteligência artificial existem algoritmos de aprendizado, notavelmente aqueles pertencentes à área de aprendizado de máquina AM , capazes de automatizar a extração do conhecimento implícito de um conjunto de dados. Dentre estes, os algoritmos de AM simbólico são aqueles que extraem um modelo de conhecimento inteligível, isto é, que pode ser facilmente interpretado pelo usuário. A utilização de AM simbólico é comum no contexto de classificação, no qual o modelo de conhecimento extraído é tal que descreve uma correlação entre um conjunto de atributos denominados premissas e um atributo particular denominado classe. Uma característica dos algoritmos de classificação é que, em geral, estes são utilizados visando principalmente a maximização das medidas de cobertura e precisão, focando a construção de um classificador genérico e preciso. Embora essa seja uma boa abordagem para automatizar processos de tomada de decisão, pode deixar a desejar quando o usuário tem o desejo de extrair um modelo de conhecimento que possa ser estudado e que possa ser útil para uma melhor compreensão do domínio. Tendo-se em vista esse cenário, o principal objetivo deste trabalho é pesquisar métodos de computação evolutiva multiobjetivo para a construção de regras de conhecimento individuais com base em critérios definidos pelo usuário. Para isso utiliza-se a biblioteca de classes e ambiente de construção de regras de conhecimento ECLE, cujo desenvolvimento remete a projetos anteriores. Outro objetivo deste trabalho consiste comparar os métodos de computação evolutiva pesquisados com métodos baseado em composição de rankings previamente existentes na ECLE. É mostrado que os métodos de computação evolutiva multiobjetivo apresentam melhores resultados que os métodos baseados em composição de rankings, tanto em termos de dominância e proximidade das soluções construídas com aquelas da fronteira Pareto-ótima quanto em termos de diversidade na fronteira de Pareto. Em otimização multiobjetivo, ambos os critérios são importantes, uma vez que o propósito da otimização multiobjetivo é fornecer não apenas uma, mas uma gama de soluções eficientes para o problema, das quais o usuário pode escolher uma ou mais soluções que apresentem os melhores compromissos entre os objetivos / Machine Learning algorithms are notable examples of Artificial Intelligence algorithms capable of automating the extraction of implicit knowledge from datasets. In particular, Symbolic Learning algorithms are those which yield an intelligible knowledge model, i.e., one which a user may easily read. The usage of Symbolic Learning is particularly common within the context of classification, which involves the extraction of knowledge such that the associated model describes correelation among a set of attributes named the premises and one specific attribute named the class. Classification algorithms usually target into creating knowledge models which maximize the measures of coverage and precision, leading to classifiers that tend to be generic and precise. Althought this constitutes a good approach to creating models that automate the decision making process, it may not yield equally good results when the user wishes to extract a knowledge model which could assist them into getting a better understanding of the domain. Having that in mind, it has been established as the main goal of this Masters thesis the research of multi-objective evolutionary computing methods to create individual knowledge rules maximizing sets of arbitrary user-defined criteria. This is achieved by employing the class library and knowledge rule construction environment ECLE, which had been developed during previous research work. A second goal of this Masters thesis is the comparison of the researched evolutionary computing methods against previously existing ranking composition methods in ECLE. It is shown in this Masters thesis that the employment of multi-objective evolutionary computing methods produces better results than those produced by the employment of ranking composition-based methods. This improvement is verified both in terms of solution dominance and proximity of the solution set to the Pareto-optimal front and in terms of Pareto-front diversity. Both criteria are important for evaluating the efficiency of multi-objective optimization algorithms, for the goal of multi-objective optimization is to provide a broad range of efficient solutions, so the user may pick one or more solutions which present the best trade-off among all objectives

Na área de inteligência artificial existem algoritmos de aprendizado, notavelmente aqueles pertencentes à área de aprendizado de máquina AM , capazes de automatizar a extração do conhecimento implícito de um conjunto de dados. Dentre estes, os algoritmos de AM simbólico são aqueles que extraem um modelo de conhecimento inteligível, isto é, que pode ser facilmente interpretado pelo usuário. A utilização de AM simbólico é comum no contexto de classificação, no qual o modelo de conhecimento extraído é tal que descreve uma correlação entre um conjunto de atributos denominados premissas e um atributo particular denominado classe. Uma característica dos algoritmos de classificação é que, em geral, estes são utilizados visando principalmente a maximização das medidas de cobertura e precisão, focando a construção de um classificador genérico e preciso. Embora essa seja uma boa abordagem para automatizar processos de tomada de decisão, pode deixar a desejar quando o usuário tem o desejo de extrair um modelo de conhecimento que possa ser estudado e que possa ser útil para uma melhor compreensão do domínio. Tendo-se em vista esse cenário, o principal objetivo deste trabalho é pesquisar métodos de computação evolutiva multiobjetivo para a construção de regras de conhecimento individuais com base em critérios definidos pelo usuário. Para isso utiliza-se a biblioteca de classes e ambiente de construção de regras de conhecimento ECLE, cujo desenvolvimento remete a projetos anteriores. Outro objetivo deste trabalho consiste comparar os métodos de computação evolutiva pesquisados com métodos baseado em composição de rankings previamente existentes na ECLE. É mostrado que os métodos de computação evolutiva multiobjetivo apresentam melhores resultados que os métodos baseados em composição de rankings, tanto em termos de dominância e proximidade das soluções construídas com aquelas da fronteira Pareto-ótima quanto em termos de diversidade na fronteira de Pareto. Em otimização multiobjetivo, ambos os critérios são importantes, uma vez que o propósito da otimização multiobjetivo é fornecer não apenas uma, mas uma gama de soluções eficientes para o problema, das quais o usuário pode escolher uma ou mais soluções que apresentem os melhores compromissos entre os objetivos / Machine Learning algorithms are notable examples of Artificial Intelligence algorithms capable of automating the extraction of implicit knowledge from datasets. In particular, Symbolic Learning algorithms are those which yield an intelligible knowledge model, i.e., one which a user may easily read. The usage of Symbolic Learning is particularly common within the context of classification, which involves the extraction of knowledge such that the associated model describes correelation among a set of attributes named the premises and one specific attribute named the class. Classification algorithms usually target into creating knowledge models which maximize the measures of coverage and precision, leading to classifiers that tend to be generic and precise. Althought this constitutes a good approach to creating models that automate the decision making process, it may not yield equally good results when the user wishes to extract a knowledge model which could assist them into getting a better understanding of the domain. Having that in mind, it has been established as the main goal of this Masters thesis the research of multi-objective evolutionary computing methods to create individual knowledge rules maximizing sets of arbitrary user-defined criteria. This is achieved by employing the class library and knowledge rule construction environment ECLE, which had been developed during previous research work. A second goal of this Masters thesis is the comparison of the researched evolutionary computing methods against previously existing ranking composition methods in ECLE. It is shown in this Masters thesis that the employment of multi-objective evolutionary computing methods produces better results than those produced by the employment of ranking composition-based methods. This improvement is verified both in terms of solution dominance and proximity of the solution set to the Pareto-optimal front and in terms of Pareto-front diversity. Both criteria are important for evaluating the efficiency of multi-objective optimization algorithms, for the goal of multi-objective optimization is to provide a broad range of efficient solutions, so the user may pick one or more solutions which present the best trade-off among all objectives

The ability to accurately predict an impending hard disk failure is important for reliable storage system design. The facility provided by most hard drive manufacturers, called S.M.A.R.T. (self-monitoring, analysis and reporting technology), has been shown by current research to have poor predictive value. The problem of finding alternatives to S.M.A.R.T. for predicting disk failure is an area of active research. In this work, we present a rule discovery methodology, and show that it is possible to construct decision support systems that can detect such failures using information recorded from live disks. It is desired that any such prediction methodology should have high accuracy and must have ease of interpretability. Black box models can deliver highly accurate solutions but do not provide an understanding of events which explains the decision given by it. To this end we explore rule based classifiers for predicting hard disk failures from various disk events. We show that it is possible to learn easy to understand rules from disk events. Our evaluation shows that our system can be tuned either to have a high failure detection rate (i.e., classify a bad disk as bad) or to have a low false alarm rate (i.e., not classify a good disk as bad). We also propose a modification of MLRules algorithm for classification of data with imbalanced class distributions. The existing algorithm, assuming relatively balanced class distributions and equal misclassfication costs, performs poorly in classification of such datasets. The performance can be considerably improved by introducing cost- sensitive learning to the existing framework.