Em muitos domínios de aplicação, a variável tempo é uma dimensão essencial. Este é o caso da robótica, na qual trajetórias de robôs podem ser interpretadas como seqüências temporais cuja ordem de ocorrência de suas componentes precisa ser considerada. Nesta dissertação, desenvolve-se um modelo de rede neural não-supervisionada para aprendizagem e reprodução de trajetórias do Robô PUMA 560. Estas trajetórias podem ter estados em comum, o que torna o processo de reprodução susceptível a ambigüidades. O modelo proposto consiste em uma rede competitiva composta por dois conjuntos de pesos sinápticos; pesos intercamadas e pesos intracamada. Pesos intercamadas conectam as unidades na camada de entrada com os neurônios da camada de saída e codificam a informação espacial contida no estímulo de entrada atual. Os pesos intracamada conectam os neurônios da camada de saída entre si, sendo divididos em dois grupos: autoconexões e conexões laterais. A função destes é codificar a ordem temporal dos estados da trajetória, estabelecendo associações entre estados consecutivos através de uma regra hebbiana. Três mecanismos adicionais são propostos de forma a tornar a aprendizagem e reprodução das trajetórias mais confiável: unidades de contexto, exclusão de neurônios e redundância na representação dos estados. A rede funciona indicando na sua saída o estado atual e o próximo estado da trajetória. As simulações com o modelo proposto ilustram a habilidade do modelo em aprender e reproduzir múltiplas trajetórias com precisão e sem ambiguidades. A rede também é capaz de reproduzir trajetórias mesmo diante de perdas de neurônios e de generalizar diante da presença de ruído nos estímulos de entrada da rede. / In many application domains, the variable time is an essential dimension. This is the case of Robotics, where robot trajectories can be interpreted as temporal sequences in which the order of occurrence of each component needs to be considered. In this dissertation, an unsupervised neural network model is developed for learning and reproducing trajectories of a Robot PUMA 560. These trajectories can have states in common, making the process of reproduction susceptible to ambiguities. The proposed model consists of a competitive network with two groups of synaptic connections: interlayer anel intralayer ones. The interlayer weights connect units in the input layer with neurons in the output layer and they encode the spatial information contained in the current input stimulus. The intralayer weights connect the neurons of the output Iayer to each other, being divided in two groups: self-connections and lateral connections. The function of these links is to encode the temporal order of the trajectory states, establishing associations among consecutive states through a Hebbian rule. Three additional mechanisms are proposed in order to make trajectory Iearning and reproduction more reliable: context units, exclusion of neurons and redundancy in the representation of the states. The model outputs the current state and the next state of the trajectory. The simulations with the proposed model illustrate the ability of the network in learning and reproducing muItiple trajectories accurateIy and without arnbiguities. In addition, the proposed neural network model is able to reproduce trajectories even when neuron failures occur and can generalize well in the presence of noise in the input stimulus.

A pesquisa em redes neurais artificiais (RNAs) está atualmente experimentando um crescente interesse por modelos que utilizem a variável tempo como um grau de liberdade extra a ser explorado nas representações neurais. Esta ênfase na codificação temporal (temporal coding) tem provocado debates inflamados nas neurociências e áreas correlatas, mas nos últimos anos o surgimento de um grande volume de dados comportamentais e fisiológicos vêm dando suporte ao papel-chave desempenhado por este tipo de representação no cérebro [BALLARD et al. (1998)]. Contribuições ao estudo da representação temporal em redes neurais vêm sendo observadas nos mais diversos tópicos de pesquisa, tais como sistemas dinâmicos não-lineares, redes oscilatórias, redes caóticas, redes com neurônios pulsantes e redes acopladas por pulsos. Como conseqüência, várias tarefas de processamento da informação têm sido investigada via codificação temporal, a saber: classificação de padrões, aprendizagem, memória associativa, controle sensório-motor, identificação de sistemas dinâmicos e robótica. Freqüentemente, porém, não fica muito claro até que ponto a modelagem dos aspectos temporais de uma tarefa contribui para aumentar a capacidade de processamento da informação de modelos neurais. Esta tese busca apresentar, de uma maneira clara e abrangente, os principais conceitos e resultados referentes à proposição de dois modelos de redes neurais não-supervisionadas (RNATs), e como estas lançam mão da codificação temporal para desempenhar melhor a tarefa que lhes é confiada. O primeiro modelo, chamado rede competitiva Hebbiana temporal (competitive temporal Hebbian - CTH), é aplicado especificamente em tarefas de aprendizagem e reprodução de trajetórias do robô manipulador PUMA 560. A rede CTH é uma rede neural competitiva cuja a principal característica é o aprendizado rápido, em apenas uma época de treinamento, de várias trajetórias complexas contendo ) elementos repetidos. As relações temporais da tarefa, representadas pela ordem temporal da trajetória, são capturadas por pesos laterais treinados via aprendizagem hebbiana. As propriedades computacionais da rede CTH são avaliadas através de simulações, bem como através da implementação de um sistema de controle distribuído para o robô PUMA 560 real. O desempenho da rede CTH é superior ao de métodos tabulares (look-up table) tradicionais de aprendizagem de trajetórias robóticas e ao de outras técnicas baseadas em redes neurais, tais como redes recorrentes supervisionadas e modelos de memória associativa bidirecional (BAM). O segundo modelo, chamado rede Auto-Organizável NARX (Self-Organizing NARX-SONARX), é baseado no conhecido algoritmo SOM, proposto por KOHONEN (1997). Do ponto de vista computacional, as propriedades de rede SONARX são avaliadas em diferentes domínios de aplicação, tais como predição de séries temporais caóticas, identificação de um atuador hidráulico e no controle preditivo de uma planta não-linear. Do ponto de vista teórico, demonstra-se que a rede SONARX pode ser utilizada como aproximador assintótico de mapeamentos dinâmicos não-lineares, graças a uma nova técnica de modelagem neural, chamada Memória Associativa Temporal via Quantização Vetorial (MATQV). A MATQV, assim como a aprendizagem hebbiana da rede CTH, é uma técnica de aprendizado associativo temporal. A rede SONARX é comparada com modelos NARX supervisionados, implementados a partir das redes MLP e RBF. Em todos os testes realizados para cada uma das tarefas citadas no parágrafo anterior, a rede SONARX tem desempenho similar ou melhor do que o apresentado por modelos supervisionados tradicionais, com um custo computacional consideravelmente menor. A rede SONARX é também comparada com a rede CTH na parendizagem de trajetórias robóticas complexas, com o intuito de destacar as principais diferenças entre os dois ) tipos de aprendizado associativo. Esta tese também propõe uma taxonomia matemática, baseada na representação por espaço de estados da teoria de sistemas, que visa classificar redes neurais não-supervisionadas temporais com ênfase em suas propriedades computacionais. Esta taxonomia tem como principal objetivo unificar a descrição de modelos neurais dinâmicos, facilitando a análise e a comparação entre diferentes arquiteturas, contrastando suas características representacionais e operacionais. Como exemplo, as redes CTH e a SONARX serão descritas usando a taxonomia proposta. / Neural network research is currently witnessing a significant shift of emphasis towards temporal coding, which uses time as an extra degree of freedom in neural representations. Temporal coding is passionately debated in neuroscience and related fields, but in the last few years a large volume of physiological and behavioral data has emerged that supports a key role for temporal coding in the brain [BALLARD et al. (1998)]. In neural networks, a great deal of research is undertaken under the topics of nonlinear dynamics, oscillatory and chaotic networks, spiking neurons, and pulse-coupled networks. Various information processing tasks are investigated using temporal coding, including pattern classification, learning, associative memory, inference, motor control, dynamical systems identification and control, and robotics. Progress has been made that substantially advances the state-of-the-art of neural computing. In many instances, however, it is unclear whether, and to what extent, the temporal aspects of the models contribute to information processing capabilities. This thesis seeks to present, in a clear and collective way, the main issues and results regarding the proposal of two unsupervised neural models, emphasizing how these networks make use of temporal coding to perform better in the task they are engaged in. The first model, called Competitive Temporal Hebbian (CTH) network, is applied specifically to learning and reproduction of trajectories of a PUMA 560 robot. The CTH model is a competitive neural network whose main characteristic is the fast learning, in just one training epoch, of multiple trajectories containing repeated elements. The temporal relationships within the task, represented by the temporal order of the elements of a given trajectory, are coded in lateral synaptic trained with hebbian learning. The computational properties of the CTH network are assessed through simulations, as well ) as through the practical implementation of a distributed control system for the real PUMA 560 robot. The CTH performs better than conventional look-up table methods for robot trajectory learning, and better than other neural-based techniques, such as supervised recurrent networks and bidirectional associative memory models. The second model, called Self-Organizing NARX (SONARX) network, is based on the well-known SOM algorithm by KOHONEN (1997). From the computational view-point, the properties of the SONARX model are evaluated in different application domains, such as prediction of chaotic time series, identification of an hydraulic actuator and predictive control of a non-linear plant. From the theoretic viewpoint, it is shown that the SONARX model can be seen as an asymptotic approximator for nonlinear dynamical mappings, thanks to a new neural modelling technique, called Vector-Quantized Temporal Associative Memory (VQTAM). This VQTAM, just like the hebbian learning rule of the CTH network, is a temporal associative memory techniques. The SONARX network is compared with supervised NARX models which based on the MLP and RBF networks. For all simulations, in each one of the forementioned application domains, the SONARX network had a similar and sometimes better performance than those observed for standard supervised models, with the additional advantage of a lower computational cost. The SONARX model is also compared with the CTH network in trajectory reproduction tasks, in order to contrast the main differences between these two types of temporal associative learning models. In this thesis, it is also proposed a mathematical taxonomy, based on the state-space representation of dynamical systems, for classification of unsupervised temporal neural networks with emphasis in their computational properties. The main goal of this taxonomy is to unify the description of dynamic neural models, ) facilitating the analysis and comparison of different architectures by constrasting their representational and operational characteristics. Is is shown how the CTH and SONARX models can be described using the proposed taxonomy.

Em muitos domínios de aplicação, a variável tempo é uma dimensão essencial. Este é o caso da robótica, na qual trajetórias de robôs podem ser interpretadas como seqüências temporais cuja ordem de ocorrência de suas componentes precisa ser considerada. Nesta dissertação, desenvolve-se um modelo de rede neural não-supervisionada para aprendizagem e reprodução de trajetórias do Robô PUMA 560. Estas trajetórias podem ter estados em comum, o que torna o processo de reprodução susceptível a ambigüidades. O modelo proposto consiste em uma rede competitiva composta por dois conjuntos de pesos sinápticos; pesos intercamadas e pesos intracamada. Pesos intercamadas conectam as unidades na camada de entrada com os neurônios da camada de saída e codificam a informação espacial contida no estímulo de entrada atual. Os pesos intracamada conectam os neurônios da camada de saída entre si, sendo divididos em dois grupos: autoconexões e conexões laterais. A função destes é codificar a ordem temporal dos estados da trajetória, estabelecendo associações entre estados consecutivos através de uma regra hebbiana. Três mecanismos adicionais são propostos de forma a tornar a aprendizagem e reprodução das trajetórias mais confiável: unidades de contexto, exclusão de neurônios e redundância na representação dos estados. A rede funciona indicando na sua saída o estado atual e o próximo estado da trajetória. As simulações com o modelo proposto ilustram a habilidade do modelo em aprender e reproduzir múltiplas trajetórias com precisão e sem ambiguidades. A rede também é capaz de reproduzir trajetórias mesmo diante de perdas de neurônios e de generalizar diante da presença de ruído nos estímulos de entrada da rede. / In many application domains, the variable time is an essential dimension. This is the case of Robotics, where robot trajectories can be interpreted as temporal sequences in which the order of occurrence of each component needs to be considered. In this dissertation, an unsupervised neural network model is developed for learning and reproducing trajectories of a Robot PUMA 560. These trajectories can have states in common, making the process of reproduction susceptible to ambiguities. The proposed model consists of a competitive network with two groups of synaptic connections: interlayer anel intralayer ones. The interlayer weights connect units in the input layer with neurons in the output layer and they encode the spatial information contained in the current input stimulus. The intralayer weights connect the neurons of the output Iayer to each other, being divided in two groups: self-connections and lateral connections. The function of these links is to encode the temporal order of the trajectory states, establishing associations among consecutive states through a Hebbian rule. Three additional mechanisms are proposed in order to make trajectory Iearning and reproduction more reliable: context units, exclusion of neurons and redundancy in the representation of the states. The model outputs the current state and the next state of the trajectory. The simulations with the proposed model illustrate the ability of the network in learning and reproducing muItiple trajectories accurateIy and without arnbiguities. In addition, the proposed neural network model is able to reproduce trajectories even when neuron failures occur and can generalize well in the presence of noise in the input stimulus.

A pesquisa em redes neurais artificiais (RNAs) está atualmente experimentando um crescente interesse por modelos que utilizem a variável tempo como um grau de liberdade extra a ser explorado nas representações neurais. Esta ênfase na codificação temporal (temporal coding) tem provocado debates inflamados nas neurociências e áreas correlatas, mas nos últimos anos o surgimento de um grande volume de dados comportamentais e fisiológicos vêm dando suporte ao papel-chave desempenhado por este tipo de representação no cérebro [BALLARD et al. (1998)]. Contribuições ao estudo da representação temporal em redes neurais vêm sendo observadas nos mais diversos tópicos de pesquisa, tais como sistemas dinâmicos não-lineares, redes oscilatórias, redes caóticas, redes com neurônios pulsantes e redes acopladas por pulsos. Como conseqüência, várias tarefas de processamento da informação têm sido investigada via codificação temporal, a saber: classificação de padrões, aprendizagem, memória associativa, controle sensório-motor, identificação de sistemas dinâmicos e robótica. Freqüentemente, porém, não fica muito claro até que ponto a modelagem dos aspectos temporais de uma tarefa contribui para aumentar a capacidade de processamento da informação de modelos neurais. Esta tese busca apresentar, de uma maneira clara e abrangente, os principais conceitos e resultados referentes à proposição de dois modelos de redes neurais não-supervisionadas (RNATs), e como estas lançam mão da codificação temporal para desempenhar melhor a tarefa que lhes é confiada. O primeiro modelo, chamado rede competitiva Hebbiana temporal (competitive temporal Hebbian - CTH), é aplicado especificamente em tarefas de aprendizagem e reprodução de trajetórias do robô manipulador PUMA 560. A rede CTH é uma rede neural competitiva cuja a principal característica é o aprendizado rápido, em apenas uma época de treinamento, de várias trajetórias complexas contendo ) elementos repetidos. As relações temporais da tarefa, representadas pela ordem temporal da trajetória, são capturadas por pesos laterais treinados via aprendizagem hebbiana. As propriedades computacionais da rede CTH são avaliadas através de simulações, bem como através da implementação de um sistema de controle distribuído para o robô PUMA 560 real. O desempenho da rede CTH é superior ao de métodos tabulares (look-up table) tradicionais de aprendizagem de trajetórias robóticas e ao de outras técnicas baseadas em redes neurais, tais como redes recorrentes supervisionadas e modelos de memória associativa bidirecional (BAM). O segundo modelo, chamado rede Auto-Organizável NARX (Self-Organizing NARX-SONARX), é baseado no conhecido algoritmo SOM, proposto por KOHONEN (1997). Do ponto de vista computacional, as propriedades de rede SONARX são avaliadas em diferentes domínios de aplicação, tais como predição de séries temporais caóticas, identificação de um atuador hidráulico e no controle preditivo de uma planta não-linear. Do ponto de vista teórico, demonstra-se que a rede SONARX pode ser utilizada como aproximador assintótico de mapeamentos dinâmicos não-lineares, graças a uma nova técnica de modelagem neural, chamada Memória Associativa Temporal via Quantização Vetorial (MATQV). A MATQV, assim como a aprendizagem hebbiana da rede CTH, é uma técnica de aprendizado associativo temporal. A rede SONARX é comparada com modelos NARX supervisionados, implementados a partir das redes MLP e RBF. Em todos os testes realizados para cada uma das tarefas citadas no parágrafo anterior, a rede SONARX tem desempenho similar ou melhor do que o apresentado por modelos supervisionados tradicionais, com um custo computacional consideravelmente menor. A rede SONARX é também comparada com a rede CTH na parendizagem de trajetórias robóticas complexas, com o intuito de destacar as principais diferenças entre os dois ) tipos de aprendizado associativo. Esta tese também propõe uma taxonomia matemática, baseada na representação por espaço de estados da teoria de sistemas, que visa classificar redes neurais não-supervisionadas temporais com ênfase em suas propriedades computacionais. Esta taxonomia tem como principal objetivo unificar a descrição de modelos neurais dinâmicos, facilitando a análise e a comparação entre diferentes arquiteturas, contrastando suas características representacionais e operacionais. Como exemplo, as redes CTH e a SONARX serão descritas usando a taxonomia proposta. / Neural network research is currently witnessing a significant shift of emphasis towards temporal coding, which uses time as an extra degree of freedom in neural representations. Temporal coding is passionately debated in neuroscience and related fields, but in the last few years a large volume of physiological and behavioral data has emerged that supports a key role for temporal coding in the brain [BALLARD et al. (1998)]. In neural networks, a great deal of research is undertaken under the topics of nonlinear dynamics, oscillatory and chaotic networks, spiking neurons, and pulse-coupled networks. Various information processing tasks are investigated using temporal coding, including pattern classification, learning, associative memory, inference, motor control, dynamical systems identification and control, and robotics. Progress has been made that substantially advances the state-of-the-art of neural computing. In many instances, however, it is unclear whether, and to what extent, the temporal aspects of the models contribute to information processing capabilities. This thesis seeks to present, in a clear and collective way, the main issues and results regarding the proposal of two unsupervised neural models, emphasizing how these networks make use of temporal coding to perform better in the task they are engaged in. The first model, called Competitive Temporal Hebbian (CTH) network, is applied specifically to learning and reproduction of trajectories of a PUMA 560 robot. The CTH model is a competitive neural network whose main characteristic is the fast learning, in just one training epoch, of multiple trajectories containing repeated elements. The temporal relationships within the task, represented by the temporal order of the elements of a given trajectory, are coded in lateral synaptic trained with hebbian learning. The computational properties of the CTH network are assessed through simulations, as well ) as through the practical implementation of a distributed control system for the real PUMA 560 robot. The CTH performs better than conventional look-up table methods for robot trajectory learning, and better than other neural-based techniques, such as supervised recurrent networks and bidirectional associative memory models. The second model, called Self-Organizing NARX (SONARX) network, is based on the well-known SOM algorithm by KOHONEN (1997). From the computational view-point, the properties of the SONARX model are evaluated in different application domains, such as prediction of chaotic time series, identification of an hydraulic actuator and predictive control of a non-linear plant. From the theoretic viewpoint, it is shown that the SONARX model can be seen as an asymptotic approximator for nonlinear dynamical mappings, thanks to a new neural modelling technique, called Vector-Quantized Temporal Associative Memory (VQTAM). This VQTAM, just like the hebbian learning rule of the CTH network, is a temporal associative memory techniques. The SONARX network is compared with supervised NARX models which based on the MLP and RBF networks. For all simulations, in each one of the forementioned application domains, the SONARX network had a similar and sometimes better performance than those observed for standard supervised models, with the additional advantage of a lower computational cost. The SONARX model is also compared with the CTH network in trajectory reproduction tasks, in order to contrast the main differences between these two types of temporal associative learning models. In this thesis, it is also proposed a mathematical taxonomy, based on the state-space representation of dynamical systems, for classification of unsupervised temporal neural networks with emphasis in their computational properties. The main goal of this taxonomy is to unify the description of dynamic neural models, ) facilitating the analysis and comparison of different architectures by constrasting their representational and operational characteristics. Is is shown how the CTH and SONARX models can be described using the proposed taxonomy.

Bovine mastitis adversely affects the dairy industry around the world. This disease is caused by a diverse range of bacteria, broadly categorised as minor and major pathogens. In-line tools that help identify these bacterial groupings in the early stages of the disease are advantageous as timely decisions could be made before the cow develops any clinical symptoms. The first objective of this research was to identify the most informative milk parameters for the detection of minor and major bacterial pathogens. The second objective of this research was to evaluate the potential of supervised and unsupervised neural network learning paradigms for the detection of minor infected and major infected quarters in the early stages of the disease. The third objective was to evaluate the effects of different proportions of infected to non-infected cases in the training data set on the correct classification rate of the supervised neural network models as there are proportionately more non-infected cases in a herd than infected cases. A database developed at Lincoln University was used to achieve the research objectives. Starting at calving, quarter milk samples were collected weekly from 112 cows for a period of fourteen weeks, resulting in 4852 samples with complete records for somatic cell count (SCC), electrical resistance, protein percentage, fat percentage, and bacteriological status. To account for the effects of the stage of lactation on milk parameters with respect to days in milking, data was divided into three days in milk ranges. In addition, cow variation was accounted for by the sire family from which the cow originated and the lactation number of each cow. Data was pre-processed before the application of advanced analytical techniques. Somatic cell score (SCS) and electrical resistance index were derived from somatic cell count and electrical resistance, respectively. After pre-processing, the data was divided into training and validation sets for the unsupervised neural network modelling experiment and, for the supervised neural network modelling experiments, the data was divided into training, calibration and validation sets. Prior to any modelling experiments, the data was analysed using statistical and multivariate visualisation techniques. Correlations (p<0.05) were found between the infection status of a quarter and its somatic cell score (SCS, 0.86), electrical resistance index (ERI, -0.59) and protein percentage (PP, 0.33). The multivariate parallel visualisation analysis validated the correlation analysis. Due to significant multicolinearity [Correlations: SCS and ERI (-0.65: p<0.05); SCS and PP (0.32: p<0.05); ERI and PP (-0.35: p<0.05)], the original variables were decorrelated using principle component analysis. SCS and ERI were found to be the most informative variables for discriminating between non-infected, minor infected and major infected cases. Unsupervised neural network (USNN) model was trained using the training data set which was extracted from the database, containing approximately equal number of randomly selected records for each bacteriological status [not infected (NI), infected with a major pathogen (MJI) and infected with a minor pathogen (MNI)]. The USNN model was validated with the remaining data using the four principle components, days in milk (DIM), lactation number (LN), sire number, and bacteriological status (BS). The specificity of the USNN model in correctly identifying non infected cases was 97%. Sensitivities for correctly detecting minor and major infections were 89% and 80%, respectively. The supervised neural network (SNN) models were trained, calibrated and validated with several sets of training, calibration and validation data, which were randomly extracted from the database in such a way that each set has a different proportion of infected to non-infected cases ranging from 1:1 to 1:10. The overall accuracy of these models based on validation data sets gradually increased with increase in the number of non-infected cases in the data sets (80% for the 1:1, 84% for 1:2, 86% for 1:4 and 93% for 1:10). Specificities of the best models for correctly recognising non-infected cases for the four data sets were 82% for 1:1, 91% for 1:2, 94% for 1:4 and 98% for 1:10. Sensitivities for correctly recognising minor infected cases for the four data sets were 86% for 1:1, 76% for 1:2, 71% for 1:4 and 44% for 1:10. Sensitivities for correctly recognising major infected cases for the four data sets were 20% for 1:1, 20% for 1:2, 30% for 1:4 and 40% for 1:10. Overall, sensitivity for the minor infected cases decreased while that of major infected cases increased with increase in the number non-infected cases in the training data set. Due to the very low prevalence of MJI category in this particular herd, results for this category may be inconclusive. This research suggests that somatic cell score and electrical resistance index of milk were the most effective variables for detecting the infection status of a quarter followed by milk protein and fat percentage. The neural network models were able to differentiate milk containing minor and major bacterial pathogens based on milk parameters associated with mastitis. It is concluded that the neural network models can be developed and incorporated into milking machines to provide an efficient and effective method for the diagnosis of mastitis.

增長層級式自我組織映射網路(GHSOM)屬於一種非監督式類神經網路，為自我組織映射網路(SOM)的延伸，擅長於對樣本分群，以輔助分析樣本族群裡的共同特徵，並且可以透過族群間存在的空間關係假設來建立分類器，進而辨別出異常的資料。 因此本研究提出一個創新的對偶方法(即為一個建立決策支援系統架構的方法)分別對舞弊與非舞弊樣本分群，首先兩類別之群組會被配對，即辨識某一特定無弊群體的非舞弊群體對照組，針對這些配對族群，套用基於不同空間假設所設立的分類規則以檢測舞弊與非舞弊群體中是否有存在某種程度的空間關係，此外並對於舞弊樣本的分群結果加入特徵萃取機制。分類績效最好的分類規則會被用來偵測受測樣本是否有舞弊的嫌疑，萃取機制的結果則會用來標示有舞弊嫌疑之受測樣本的舞弊行為特徵以及相關的輸入變數，以做為後續的決策輔助。 更明確地說，本研究分別透過非舞弊樣本與舞弊樣本建立一個非舞弊GHSOM樹以及舞弊GHSOM樹，且針對每一對GHSOM群組建立分類規則，其相應的非舞弊/舞弊為中心規則會適應性地依循決策者的風險偏好最佳化調整規則界線，整體而言較優的規則會被決定為分類規則。非舞弊為中心的規則象徵絕大多數的舞弊樣本傾向分布於非舞弊樣本的周圍，而舞弊為中心的規則象徵絕大多數的非舞弊樣本傾向分布於舞弊樣本的周圍。 此外本研究加入了一個特徵萃取機制來發掘舞弊樣本分群結果中各群組之樣本資料的共同特質，其包含輸入變數的特徵以及舞弊行為模式，這些資訊將能輔助決策者(如資本提供者)評估受測樣本的誠實性，輔助決策者從分析結果裡做出更進一步的分析來達到審慎的信用決策。 本研究將所提出的方法套用至財報舞弊領域(屬於財務舞弊偵測的子領域)進行實證，實驗結果證實樣本之間存在特定的空間關係，且相較於其他方法如SVM、SOM+LDA和GHSOM+LDA皆具有更佳的分類績效。因此顯示本研究所提出的機制可輔助驗證財務相關數據的可靠性。此外，根據SOM的特質，即任何受測樣本歸類到某特定族群時，該族群訓練樣本的舞弊行為特徵將可以代表此受測樣本的特徵推論。這樣的原則可以用來協助判斷受測樣本的可靠性，並可供持續累積成一個舞弊知識庫，做為進一步分析以及制定相關信用決策的參考。本研究所提出之基於對偶方法的決策支援系統架構可以被套用到其他使用財務數據為資料來源的財務舞弊偵測情境中，作為輔助決策的基礎。 / The Growing Hierarchical Self-Organizing Map (GHSOM) is extended from the Self-Organizing Map (SOM). The GHSOM’s unsupervised learning nature such as the adaptive group size as well as the hierarchy structure renders its availability to discover the statistical salient features from the clustered groups, and could be used to set up a classifier for distinguishing abnormal data from regular ones based on spatial relationships between them. Therefore, this study utilizes the advantage of the GHSOM and pioneers a novel dual approach (i.e., a proposal of a DSS architecture) with two GHSOMs, which starts from identifying the counterparts within the clustered groups. Then, the classification rules are formed based on a certain spatial hypothesis, and a feature extraction mechanism is applied to extract features from the fraud clustered groups. The dominant classification rule is adapted to identify suspected samples, and the results of feature extraction mechanism are used to pinpoint their relevant input variables and potential fraud activities for further decision aid. Specifically, for the financial fraud detection (FFD) domain, a non-fraud (fraud) GHSOM tree is constructed via clustering the non-fraud (fraud) samples, and a non-fraud-central (fraud-central) rule is then tuned via inputting all the training samples to determine the optimal discrimination boundary within each leaf node of the non-fraud (fraud) GHSOM tree. The optimization renders an adjustable and effective rule for classifying fraud and non-fraud samples. Following the implementation of the DSS architecture based on the proposed dual approach, the decision makers can objectively set their weightings of type I and type II errors. The classification rule that dominates another is adopted for analyzing samples. The dominance of the non-fraud-central rule leads to an implication that most of fraud samples cluster around the non-fraud counterpart, meanwhile the dominance of fraud-central rule leads to an implication that most of non-fraud samples cluster around the fraud counterpart. Besides, a feature extraction mechanism is developed to uncover the regularity of input variables and fraud categories based on the training samples of each leaf node of a fraud GHSOM tree. The feature extraction mechanism involves extracting the variable features and fraud patterns to explore the characteristics of fraud samples within the same leaf node. Thus can help decision makers such as the capital providers evaluate the integrity of the investigated samples, and facilitate further analysis to reach prudent credit decisions. The experimental results of detecting fraudulent financial reporting (FFR), a sub-field of FFD, confirm the spatial relationship among fraud and non-fraud samples. The outcomes given by the implemented DSS architecture based on the proposed dual approach have better classification performance than the SVM, SOM+LDA, GHSOM+LDA, SOM, BPNN and DT methods, and therefore show its applicability to evaluate the reliability of the financial numbers based decisions. Besides, following the SOM theories, the extracted relevant input variables and the fraud categories from the GHSOM are applicable to all samples classified into the same leaf nodes. This principle makes that the extracted pre-warning signal can be applied to assess the reliability of the investigated samples and to form a knowledge base for further analysis to reach a prudent decision. The DSS architecture based on the proposed dual approach could be applied to other FFD scenarios that rely on financial numbers as a basis for decision making.