In this paper we describe a technique of classifier combination used in a human identification system. The system integrates all available features from multi-modal sources within a Bayesian framework. The framework allows representinga class of popular classifier combination rules and methods within a single formalism. It relies on a “per-class” measure of confidence derived from performance of each classifier on training data that is shown to improve performance on a synthetic data set. The method is especially relevant in autonomous surveillance setting where varying time scales and missing features are a common occurrence. We show an application of this technique to the real-world surveillance database of video and audio recordings of people collected over several weeks in the office setting.

Access technologies often deal with the classification of several physiological and biomechanical signals. In most previous studies involving access technologies, a single classifier has been trained. Despite reported success of these single classifiers, classification accuracies are often below clinically viable levels. One approach to improve upon the performance of these classifiers is to utilize the state of- the-art multiple classifier systems (MCS). Because MCS invoke more than one classifier, more information can be exploited from the signals, potentially leading to higher classification performance than that achievable with single classifiers. Moreover, by decreasing the feature space dimensionality of each classifier, the speed of the system can be increased. MCSs may combine classifiers on three levels: abstract, rank, or measurement level. Among them, abstract-level MCSs have been the most widely applied in the literature given the flexibility of the abstract level output, i.e., class labels may be derived from any type of classifier and outputs from multiple classifiers, each designed within a different context, can be easily combined. In this thesis, we develop two new abstract-level MCSs based on "reputation" values of individual classifiers: the static reputation-based algorithm (SRB) and the dynamic reputation-based algorithm (DRB). In SRB, each individual classifier is applied to a “validation set”, which is disjoint from training and test sets, to estimate its reputation value. Then, each individual classifier is assigned a weight proportional to its reputation value. Finally, the total decision of the classification system is computed using Bayes rule. We have applied this method to the problem of dysphagia detection in adults with neurogenic swallowing difficulties. The aim was to discriminate between safe and unsafe swallows. The weighted classification accuracy exceeded 85% and, because of its high sensitivity, the SRB approach was deemed suitable for screening purposes. In the next step of this dissertation, I analyzed the SRB algorithm mathematically and examined its asymptotic behavior. Specifically, I contrasted the SRB performance against that of majority voting, the benchmark abstract-level MCS, in the presence of different types of noise. In the second phase of this thesis, I exploited the idea of the Dirichlet reputation system to develop a new MCS method, the dynamic reputation-based algorithm, which is suitable for the classification of non-stationary signals. In this method, the reputation of each classifier is updated dynamically whenever a new sample is classified. At any point in time, a classifier’s reputation reflects the classifier’s performance on both the validation and the test sets. Therefore, the effect of random high-performance of weak classifiers is appropriately moderated and likewise, the effect of a poorly performing individual classifier is mitigated as its reputation value, and hence overall influence on the final decision is diminished. We applied DRB to the challenging problem of discerning physiological responses from nonverbal youth with severe disabilities. The promising experimental results encourage further development of reputation-based multi-classifier systems in the domain of access technology research.

Access technologies often deal with the classification of several physiological and biomechanical signals. In most previous studies involving access technologies, a single classifier has been trained. Despite reported success of these single classifiers, classification accuracies are often below clinically viable levels. One approach to improve upon the performance of these classifiers is to utilize the state of- the-art multiple classifier systems (MCS). Because MCS invoke more than one classifier, more information can be exploited from the signals, potentially leading to higher classification performance than that achievable with single classifiers. Moreover, by decreasing the feature space dimensionality of each classifier, the speed of the system can be increased. MCSs may combine classifiers on three levels: abstract, rank, or measurement level. Among them, abstract-level MCSs have been the most widely applied in the literature given the flexibility of the abstract level output, i.e., class labels may be derived from any type of classifier and outputs from multiple classifiers, each designed within a different context, can be easily combined. In this thesis, we develop two new abstract-level MCSs based on "reputation" values of individual classifiers: the static reputation-based algorithm (SRB) and the dynamic reputation-based algorithm (DRB). In SRB, each individual classifier is applied to a “validation set”, which is disjoint from training and test sets, to estimate its reputation value. Then, each individual classifier is assigned a weight proportional to its reputation value. Finally, the total decision of the classification system is computed using Bayes rule. We have applied this method to the problem of dysphagia detection in adults with neurogenic swallowing difficulties. The aim was to discriminate between safe and unsafe swallows. The weighted classification accuracy exceeded 85% and, because of its high sensitivity, the SRB approach was deemed suitable for screening purposes. In the next step of this dissertation, I analyzed the SRB algorithm mathematically and examined its asymptotic behavior. Specifically, I contrasted the SRB performance against that of majority voting, the benchmark abstract-level MCS, in the presence of different types of noise. In the second phase of this thesis, I exploited the idea of the Dirichlet reputation system to develop a new MCS method, the dynamic reputation-based algorithm, which is suitable for the classification of non-stationary signals. In this method, the reputation of each classifier is updated dynamically whenever a new sample is classified. At any point in time, a classifier’s reputation reflects the classifier’s performance on both the validation and the test sets. Therefore, the effect of random high-performance of weak classifiers is appropriately moderated and likewise, the effect of a poorly performing individual classifier is mitigated as its reputation value, and hence overall influence on the final decision is diminished. We applied DRB to the challenging problem of discerning physiological responses from nonverbal youth with severe disabilities. The promising experimental results encourage further development of reputation-based multi-classifier systems in the domain of access technology research.

To solve the challenging pattern classification problem, machine learning researchers have extensively studied Multiple Classifier Systems (MCSs). The motivations for combining classifiers are found in the literature from the statistical, computational and representational perspectives. Although the results of classifier combination does not always outperform the best individual classifier in the ensemble, empirical studies have demonstrated its superiority for various applications. A number of viable methods to design MCSs have been developed including bagging, adaboost, rotation forest, and random subspace. They have been successfully applied to solve various tasks. Currently, most of the research is being conducted on the behavior patterns of the base classifiers in the ensemble. However, a discussion from the learning point of view may provide insights into the robust design of MCSs. In this thesis, Generalized Exhaustive Search and Aggregation (GESA) method is developed for this objective. Robust performance is achieved using GESA by dynamically adjusting the trade-off between fitting the training data adequately and preventing the overfitting problem. Besides its learning algorithm, GESA is also distinguished from traditional designs by its architecture and level of decision-making. GESA generates a collection of ensembles and dynamically selects the most appropriate ensemble for decision-making at the local level. Although GESA provides a good improvement over traditional approaches, it is not very data-adaptive. A data- adaptive design of MCSs demands that the system can adaptively select representations and classifiers to generate effective decisions for aggregation. Another weakness of GESA is its high computation cost which prevents it from being scaled to large ensembles. Generalized Adaptive Ensemble Generation and Aggregation (GAEGA) is an extension of GESA to overcome these two difficulties. GAEGA employs a greedy algorithm to adaptively select the most effective representations and classifiers while excluding the noise ones as much as possible. Consequently, GAEGA can generate fewer ensembles and significantly reduce the computation cost. Bootstrapped Adaptive Ensemble Generation and Aggregation (BAEGA) is another extension of GESA, which is similar with GAEGA in the ensemble generation and decision aggregation. BAEGA adopts a different data manipulation strategy to improve the diversity of the generated ensembles and utilize the information in the data more effectively. As a specific application, the classification of time series data is chosen for the research reported in this thesis. This type of data contains dynamic information and proves to be more complex than others. Multiple Input Representation-Adaptive Ensemble Generation and Aggregation (MIR-AEGA) is derived from GAEGA for the classification of time series data. MIR-AEGA involves some novel representation methods that proved to be effective for time series data. All the proposed methods including GESA, GAEGA, MIR-AEGA, and BAEGA are tested on simulated and benchmark data sets from popular data repositories. The experimental results confirm that the newly developed methods are effective and efficient.

To solve the challenging pattern classification problem, machine learning researchers have extensively studied Multiple Classifier Systems (MCSs). The motivations for combining classifiers are found in the literature from the statistical, computational and representational perspectives. Although the results of classifier combination does not always outperform the best individual classifier in the ensemble, empirical studies have demonstrated its superiority for various applications. A number of viable methods to design MCSs have been developed including bagging, adaboost, rotation forest, and random subspace. They have been successfully applied to solve various tasks. Currently, most of the research is being conducted on the behavior patterns of the base classifiers in the ensemble. However, a discussion from the learning point of view may provide insights into the robust design of MCSs. In this thesis, Generalized Exhaustive Search and Aggregation (GESA) method is developed for this objective. Robust performance is achieved using GESA by dynamically adjusting the trade-off between fitting the training data adequately and preventing the overfitting problem. Besides its learning algorithm, GESA is also distinguished from traditional designs by its architecture and level of decision-making. GESA generates a collection of ensembles and dynamically selects the most appropriate ensemble for decision-making at the local level. Although GESA provides a good improvement over traditional approaches, it is not very data-adaptive. A data- adaptive design of MCSs demands that the system can adaptively select representations and classifiers to generate effective decisions for aggregation. Another weakness of GESA is its high computation cost which prevents it from being scaled to large ensembles. Generalized Adaptive Ensemble Generation and Aggregation (GAEGA) is an extension of GESA to overcome these two difficulties. GAEGA employs a greedy algorithm to adaptively select the most effective representations and classifiers while excluding the noise ones as much as possible. Consequently, GAEGA can generate fewer ensembles and significantly reduce the computation cost. Bootstrapped Adaptive Ensemble Generation and Aggregation (BAEGA) is another extension of GESA, which is similar with GAEGA in the ensemble generation and decision aggregation. BAEGA adopts a different data manipulation strategy to improve the diversity of the generated ensembles and utilize the information in the data more effectively. As a specific application, the classification of time series data is chosen for the research reported in this thesis. This type of data contains dynamic information and proves to be more complex than others. Multiple Input Representation-Adaptive Ensemble Generation and Aggregation (MIR-AEGA) is derived from GAEGA for the classification of time series data. MIR-AEGA involves some novel representation methods that proved to be effective for time series data. All the proposed methods including GESA, GAEGA, MIR-AEGA, and BAEGA are tested on simulated and benchmark data sets from popular data repositories. The experimental results confirm that the newly developed methods are effective and efficient.

Most software programs that deal with mathematical objects require input expressions to be linearized using somewhat awkward and unfamiliar string-based syntax. It is natural to desire a method for inputting mathematics using the same two-dimensional syntax employed with pen and paper, and the increasing prevalence of pen- and touch-based interfaces causes this topic to be of practical as well as theoretical interest. Accurately recognizing two-dimensional mathematical notation is a difficult problem that requires not only theoretical advancement over the traditional theories of string-based languages, but also careful consideration of runtime efficiency, data organization, and other practical concerns that arise during system construction. This thesis describes the math recognizer used in the MathBrush pen-math system. At a high level, the two-dimensional syntax of mathematical writing is formalized using a relational grammar. Rather than reporting a single recognition result, all recognizable interpretations of the input are simultaneously represented in a data structure called a parse forest. Individual interpretations may be extracted from the forest and reported one by one as the user requests them. These parsing techniques necessitate robust tree scoring functions, which themselves rely on several lower-level recognition processes for stroke grouping, symbol recognition, and spatial relation classification. The thesis covers the recognition, parsing, and scoring aspects of the MathBrush recognizer, as well as the algorithms and assumptions necessary to combine those systems and formalisms together into a useful and efficient software system. The effectiveness of the resulting system is measured through two accuracy evaluations. One evaluation uses a novel metric based on user effort, while the other replicates the evaluation process of an international accuracy competition. The evaluations show that not only is the performance of the MathBrush recognizer improving over time, but it is also significantly more accurate than other academic recognition systems.

El aprendizaje autom�atico es un �area de investigaci�on que proporciona algoritmos y t�ecnicas que son capaces de aprender autom�aticamente a partir de experiencias pasadas. Estas t�ecnicas son esenciales en el �area de descubrimiento de conocimiento de bases de datos (KDD), cuya fase principal es t�ÿpicamente conocida como miner�ÿa de datos. El proceso de KDD se puede ver como el aprendizaje de un modelo a partir de datos anteriores (generaci�on del modelo) y la aplicaci�on de este modelo a nuevos datos (utilizaci�on del modelo). La fase de utilizaci�on del modelo es muy importante, porque los usuarios y, muy especialmente, las organizaciones toman las decisiones dependiendo del resultado de los modelos. Por lo general, cada modelo se aprende de forma independiente, intentando obtener el mejor resultado (local). Sin embargo, cuando varios modelos se usan conjuntamente, algunos de ellos pueden depender los unos de los otros (por ejemplo, las salidas de un modelo pueden ser las entradas de otro) y aparecen restricciones. En este escenario, la mejor decisi�on local para cada problema tratado individualmente podr�ÿa no dar el mejor resultado global, o el resultado obtenido podr�ÿa no ser v�alido si no cumple las restricciones del problema. El �area de administraci�on de la relaci�on con los clientes (CRM) ha dado origen a problemas reales donde la miner�ÿa de datos y la optimizaci�on (global) deben ser usadas conjuntamente. Por ejemplo, los problemas de prescripci�on de productos tratan de distinguir u ordenar los productos que ser�an ofrecidos a cada cliente (o sim�etricamente, elegir los clientes a los que se les deber�ÿa de ofrecer los productos). Estas �areas (KDD, CRM) carecen de herramientas para tener una visi�on m�as completa de los problemas y una mejor integraci�on de los modelos de acuerdo a sus interdependencias y las restricciones globales y locales. La aplicaci�on cl�asica de miner�ÿa de datos a problemas de prescripci�on de productos, por lo general, ha / Bella Sanjuán, A. (2012). Model Integration in Data Mining: From Local to Global Decisions [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/16964 / Palancia

Thesis (MSc)--Stellenbosch University, 2011. / ENGLISH ABSTRACT: An off-line signature verification system attempts to authenticate the identity of an individual by examining his/her handwritten signature, after it has been successfully extracted from, for example, a cheque, a debit or credit card transaction slip, or any other legal document. The questioned signature is typically compared to a model trained from known positive samples, after which the system attempts to label said signature as genuine or fraudulent. Classifier fusion is the process of combining individual classifiers, in order to construct a single classifier that is more accurate, albeit computationally more complex, than its constituent parts. A combined classifier therefore consists of an ensemble of base classifiers that are combined using a specific fusion strategy. In this dissertation a novel off-line signature verification system, using a multi-hypothesis approach and classifier fusion, is proposed. Each base classifier is constructed from a hidden Markov model (HMM) that is trained from features extracted from local regions of the signature (local features), as well as from the signature as a whole (global features). To achieve this, each signature is zoned into a number of overlapping circular retinas, from which said features are extracted by implementing the discrete Radon transform. A global retina, that encompasses the entire signature, is also considered. Since the proposed system attempts to detect high-quality (skilled) forgeries, it is unreasonable to assume that samples of these forgeries will be available for each new writer (client) enrolled into the system. The system is therefore constrained in the sense that only positive training samples, obtained from each writer during enrolment, are available. It is however reasonable to assume that both positive and negative samples are available for a representative subset of so-called guinea-pig writers (for example, bank employees). These signatures constitute a convenient optimisation set that is used to select the most proficient ensemble. A signature, that is claimed to belong to a legitimate client (member of the general public), is therefore rejected or accepted based on the majority vote decision of the base classifiers within the most proficient ensemble. When evaluated on a data set containing high-quality imitations, the inclusion of local features, together with classifier combination, significantly increases system performance. An equal error rate of 8.6% is achieved, which compares favorably to an achieved equal error rate of 12.9% (an improvement of 33.3%) when only global features are considered. Since there is no standard international off-line signature verification data set available, most systems proposed in the literature are evaluated on data sets that differ from the one employed in this dissertation. A direct comparison of results is therefore not possible. However, since the proposed system utilises significantly different features and/or modelling techniques than those employed in the above-mentioned systems, it is very likely that a superior combined system can be obtained by combining the proposed system with any of the aforementioned systems. Furthermore, when evaluated on the same data set, the proposed system is shown to be significantly superior to three other systems recently proposed in the literature. / AFRIKAANSE OPSOMMING: Die doel van ’n statiese handtekening-verifikasiestelsel is om die identiteit van ’n individu te bekragtig deur sy/haar handgeskrewe handtekening te analiseer, nadat dit suksesvol vanaf byvoorbeeld ’n tjek,’n debiet- of kredietkaattransaksiestrokie, of enige ander wettige dokument onttrek is. Die bevraagtekende handtekening word tipies vergelyk met ’n model wat afgerig is met bekende positiewe voorbeelde, waarna die stelsel poog om die handtekening as eg of vervals te klassifiseer. Klassifiseerder-fusie is die proses waardeer individuele klassifiseerders gekombineer word, ten einde ’n enkele klassifiseerder te konstrueer, wat meer akkuraat, maar meer berekeningsintensief as sy samestellende dele is. ’n Gekombineerde klassifiseerder bestaan derhalwe uit ’n ensemble van basis-klassifiseerders, wat gekombineer word met behulp van ’n spesifieke fusie-strategie. In hierdie projek word ’n nuwe statiese handtekening-verifikasiestelsel, wat van ’n multi-hipotese benadering en klassifiseerder-fusie gebruik maak, voorgestel. Elke basis-klassifiseerder word vanuit ’n verskuilde Markov-model (HMM) gekonstrueer, wat afgerig word met kenmerke wat vanuit lokale gebiede in die handtekening (lokale kenmerke), sowel as vanuit die handtekening in geheel (globale kenmerke), onttrek is. Ten einde dit te bewerkstellig, word elke handtekening in ’n aantal oorvleulende sirkulêre retinas gesoneer, waaruit kenmerke onttrek word deur die diskrete Radon-transform te implementeer. ’n Globale retina, wat die hele handtekening in beslag neem, word ook beskou. Aangesien die voorgestelde stelsel poog om hoë-kwaliteit vervalsings op te spoor, is dit onredelik om te verwag dat voorbeelde van hierdie handtekeninge beskikbaar sal wees vir elke nuwe skrywer (kliënt) wat vir die stelsel registreer. Die stelsel is derhalwe beperk in die sin dat slegs positiewe afrigvoorbeelde, wat bekom is van elke skrywer tydens registrasie, beskikbaar is. Dit is egter redelik om aan te neem dat beide positiewe en negatiewe voorbeelde beskikbaar sal wees vir ’n verteenwoordigende subversameling van sogenaamde proefkonynskrywers, byvoorbeeld bankpersoneel. Hierdie handtekeninge verteenwoordig ’n gereieflike optimeringstel, wat gebruik kan word om die mees bekwame ensemble te selekteer. ’n Handtekening, wat na bewering aan ’n wettige kliënt (lid van die algemene publiek) behoort, word dus verwerp of aanvaar op grond van die meerderheidstem-besluit van die basis-klassifiseerders in die mees bekwame ensemble. Wanneer die voorgestelde stelsel op ’n datastel, wat hoë-kwaliteit vervalsings bevat, ge-evalueer word, verhoog die insluiting van lokale kenmerke en klassifiseerder-fusie die prestasie van die stelsel beduidend. ’n Gelyke foutkoers van 8.6% word behaal, wat gunstig vergelyk met ’n gelyke foutkoers van 12.9% (’n verbetering van 33.3%) wanneer slegs globale kenmerke gebruik word. Aangesien daar geen standard internasionale statiese handtekening-verifikasiestelsel bestaan nie, word die meeste stelsels, wat in die literatuur voorgestel word, op ander datastelle ge-evalueer as die datastel wat in dié projek gebruik word. ’n Direkte vergelyking van resultate is dus nie moontlik nie. Desnieteenstaande, aangesien die voorgestelde stelsel beduidend ander kenmerke en/of modeleringstegnieke as dié wat in bogenoemde stelsels ingespan word gebruik, is dit hoogs waarskynlik dat ’n superieure gekombineerde stelsel verkry kan word deur die voorgestelde stelsel met enige van bogenoemde stelsels te kombineer. Voorts word aangetoon dat, wanneer op dieselfde datastel geevalueerword, die voorgestelde stelstel beduidend beter vaar as drie ander stelsels wat onlangs in die literatuur voorgestel is.

Recognizing handwritten mathematical content is a challenging problem, and more so when such content appears in classroom videos. However, given the fact that in such videos the handwritten text and the accompanying audio refer to the same content, a combination of video and audio based recognizer has the potential to significantly improve the content recognition accuracy. This dissertation, using a combination of video and audio based recognizers, focuses on improving the recognition accuracy associated with handwritten mathematical content in such videos. Our approach makes use of a video recognizer as the primary recognizer and a multi-stage assembly, developed as part of this research, is used to facilitate effective combination with an audio recognizer. Specifically, we address the following challenges related to audio-video based handwritten mathematical content recognition: (1) Video Preprocessing - generates a timestamped sequence of segmented characters from the classroom video in the face of occlusions and shadows caused by the instructor, (2) Ambiguity Detection - determines the subset of input characters that may have been incorrectly recognized by the video based recognizer and forwards this subset for disambiguation, (3) A/V Synchronization - establishes correspondence between the handwritten character and the spoken content, (4) A/V Combination - combines the synchronized outputs from the video and audio based recognizers and generates the final recognized character, and (5) Grammar Assisted A/V Based Mathematical Content Recognition - utilizes a base mathematical speech grammar for both character and structure disambiguation. Experiments conducted using videos recorded in a classroom-like environment demonstrate the significant improvements in recognition accuracy that can be achieved using our techniques.

[pt] Esta dissertação investigou a criação de comitês de classificadores baseados em Redes Neurais Multilayer Perceptron (Redes MLP, abreviadamente). Isso foi feito em dois passos: primeiro, aplicando-se procedimentos para criação de redes complementares, i.e, redes individualmente eficazes mas que cometem erros distintos; segundo, testando- se sobre essas redes alguns dos principais métodos de combinação disponíveis. Dentre os procedimentos para criação de redes complementares, foi dado enfoque para os baseados em alteração do conjunto de treinamento. Os métodos Bootstrap e Arc-x4 foram escolhidos para serem utilizados no estudo de casos, juntamente com o método RDP (Replicação Dirigida de Padrões). No que diz respeito aos métodos de combinação disponíveis, foi dada particular atenção ao método de combinação por integrais nebulosas. Além deste método, implementou-se combinação por média, votação por pluralidade e Borda count. As aplicações escolhidas para teste envolveram duas vertentes importantes na área de visão computacional - Classificação de Coberturas de Solo por Imagens de Satélite e Reconhecimento de Expressões Faciais. Embora ambas pertençam à mesma área de conhecimento, foram escolhidas de modo a representar níveis de dificuldade diferentes como tarefas de classificação - enquanto a primeira contou com um grande número de padrões disponíveis, a segunda foi comparativamente limitada nesse sentido. Como resultado final, comprovou-se a viabilidade da utilização de comitês em problemas de classificação, mesmo com as possíveis variações de desempenho relacionadas com a complexidade desses problemas. O método de combinação baseado em integrais nebulosas mostrou-se particularmente eficiente quando associado ao procedimento RDP para formação das redes comissionadas, mas nem sempre foi satisfatório. Considerado individualmente, o RDP tem a limitação de criar, no máximo, tantas redes quanto forem as classes consideradas em um problema; porém, quando este número de redes foi considerado como base de comparação, o RDP se mostrou, na média de todos os métodos de combinação testados, mais eficaz que os procedimentos Bootstrap e Arc-x4. Por outro lado, tanto o Bootstrap quanto o Arc-x4 têm a importante vantagem de permitirem a formação de um número crescente de membros, o que quase sempre acarretou em melhorias de desempenho global em relação ao RDP. / [en] The present dissertation investigated the creation of classifier committees based on Multilayer Perceptron Neural Networks (MLP Networks, for short). This was done in two parts: first, by applying procedures for creating complementary networks, i.e., networks that are individually accurate but cause distinct misclassifications; second, by assessing different combining methods to these network`s outputs. Among the procedures for creating committees members, the main focus was set to the ones based on changes to the training set . Bootstrap and Arc-x4 were chosen to be used at the experiments, along with the RDP procedure (translated as Driven Pattern Replication). With respect to the available combining methods, special attention was paid to fuzzy integrals combination. Average combination, plurality voting and Borda count were also implemented. The chosen experimental applications included interesting branches from computer vision: Land Cover Classification from Satellite Images and Facial Expression Recognition. These applications were specially interesting, in the sense they represent two different levels of difficulty as classification tasks - while the first had a great number of available patterns, the second was comparatively limited in this way. This work proved the viability of using committees in classification problems, despite the small performance fluctuations related to these problems complexity. The fuzzy integrals method has shown to be particularly interesting when coupled with the RDP procedure for committee creation, but was not always satisfactory. Taken alone, the RDP has the limitation of creating, at most, as many networks as there are classes to be considered at the problem at hand; however, when this number of networks was considered as the basis for comparison, this procedure outperformed, taking into account average combining results, both Bootstrap and Arc- x4. On the other hand, these later procedures have the important advantage of allowing the creation of an increasing number of committee members, what almost always increased global performance in comparison to RDP. / [es] Esta disertación investigó la creación de comités de clasificadores basados en Redes Neurales Multilayer Perceptron (Redes MLP, abreviadamente). Esto fue ejecutado en dos pasos: primeiro, aplicando procedimentos para la creación de redes complementares, esto es, redes que individualmente son eficaces pero que cometen erros diferentes; segundo, probando sobre esas redes, algunos de los principales métodos de combinación disponibles. Dentro de los procedimentos para la creación de redes complementares, se eligieron los basados en alteración del conjunto de entrenamiento. Los métodos Bootstrap y Arc-x4 fueron seleccionados para utilizarlos em el estudio de casos, conjuntamente con el método RDP (Replicación Dirigida de Padrones). Con respecto a los métodos de combinación disponibles, se le dió particular atención al método de combinación por integrales nebulosas. Además de este método, se implementaron: combinación por media, votación por pluralidad y Borda cont. Las aplicaciones seleccionadas para pruebas consideran dos vertientes importantes en la área de visión computacional - Clasificación de Coberturas de Suelo por Imágenes de Shastalite y Reconocimiento de Expresiones Faciales. Aunque ambas pertencen a la misma área de conocimento, fueron seleccionadas de modo con diferentes níveles de dificuldad como tareas de clasificación - Mientras la primera contó con un gran número de padrones disponibles, la segunda fue comparativamente limitada em ese sentido. Como resultado final, se comprobó la viabilidad de la utilización de comités en problemas de clasificación, incluso con las posibles variaciones de desempeño relacionadas con la complejidad de esos problemas. El método de combinación basado en integrales nebulosas se mostró particularmente eficiente asociado al procedimiento RDP para formación de las redes comisionadas, pero no siempre fue satisfactorio. Considerado individualmente, el RDP tiene la limitación de crear, como máximo, tantas redes como clases consideradas en un problema; sin embargo, cuando el número de redes fue considerado como base de comparación, el RDP se mostró más eficaz, en la media de todos los métodos de combinación, que los procedimentos Bootstrap y Arc-x4. Por otro lado, tanto el Bootstrap como el Arc-x4 tiene la importante ventaja de permitir la formación de un número cresciente de miembros, lo que generalmente mejora el desempeño global en relación al RDP.