Thesis (Ph.D.)--Hong Kong University of Science and Technology, 2007. / Includes bibliographical references (leaves 106-113). Also available in electronic version.

Thesis (MEng)--Stellenbosch University, 2014. / ENGLISH ABSTRACT: The cross-entropy method for multi-objective optimisation (MOO CEM) was recently introduced by Bekker & Aldrich (2010) and Bekker (2012). Results presented by both show great promise. The MOO CEM assumes that decision variables are independent. As a consequence, the question arises: under which circumstances would an algorithm that accounts for relationships between decision variables outperform the MOO CEM? Two algorithms reported to account for relationships between decision variables, the multi-objective covariance matrix adaptation evolution strategy (MOCMA- ES) and Pareto di erential evolution (PDE), are selected for comparison. In addition, two hybrid algorithms (Hybrid 1 and Hybrid 2) based on the MOO CEM are created. These ve algorithms are applied to a set of 46 continuous problems, six instances of the mission-ready resource (MRR) problem, and three instances of a dynamic, stochastic bu er allocation problem (BAP). Performance is measured using the hypervolume indicator and Mann-Whitney U-tests. One of the primary ndings is that accounting for relationships between decision variables is bene cial when solving small to medium-sized problems. In these cases, the MO-CMA-ES typically outperforms the other algorithms. However, on large problems, Hybrid 1 and the MOO CEM typically perform best. / AFRIKAANSE OPSOMMING: Die kruis-entropie metode vir meerdoelige optimering (MOO CEM) is onlangs deur Bekker & Aldrich (2010) en Bekker (2012) bekendgestel. Hul resultate is belowend. Die MOO CEM neem aan dat besluitnemingsveranderlikes onafhanklik is van mekaar. Gevolglik ontstaan die vraag: onder watter omstandighede sal 'n optimeringsalgoritme wat moontlike verhoudings tussen besluitnemingsveranderlikes in ag neem, beter vaar as die MOO CEM? Twee bestaande algoritmes, beide gerapporteer vir hul vermo e om moontlike verhoudings tussen besluitnemingsveranderlikes in ag te neem, naamlik die meerdoelige optimering kovariansiematriksaanpassing-evolusiestrategie (MO-CMA-ES) en Pareto afgeleide evolusie (PDE), word met die MOO CEM vergelyk. Twee nuwe hibriedalgoritmes (Hibried 1 en Hibried 2) word ook ter wille van di e vergelyking geskep. Die vyf algoritmes word op 'n stel van 46 kontinue probleme, ses statiese kombinatoriese gevalle en drie dinamies, stogastiese gevalle toegepas. Die prestasie van die algoritmes word deur middel van die hipervolume-aanwyser en Mann-Whitney U-toetse gemeet. 'n Prim^ere bevinding is dat dit voordelig is om moontlike verhoudings tussen besluitnemingsveranderlikes in ag te neem wanneer klein na medium-grootte probleme opgelos word. Vir hierdie gevalle presteer die MO-CMA-ES tipies beter as die ander algoritmes. Vir groot probleme presteer Hibried 1 en die MOO CEM beter as die ander algoritmes. / National Research Foundation

Thesis (MSc (Mathematical Sciences. Computer Science))--University of Stellenbosch, 2007. / It is widely recognized that Traffic engineering (TE) mechanisms have to be added to the IP transport functionalities to provide QoS guarantees while ensuring efficient use of network resources. Traffic engineering is a network management technique which routes traffic to where bandwidth is available in the network to achieve QoS agreements between current and future demands and the available network resources. Multi-path routing has been proven to be a more efficient TE mechanism than Shortest Path First (SPF) routing in terms of proffit maximization and resource usage optimization. However the identiffication of set of paths over which traffic is forwarded from source to the destination and the distribution of traffic among these paths are two issues that have been widely addressed by the IP community but remain an open issue for the emerging generation IP networks. Building upon different frameworks, this thesis revisits the issue of multi-path routing to present and evaluate the performance of different traffic splitting mechanisms to achieve QoS routing in Multi-Protocol Label Switching (MPLS) and Wireless Sensor Networks (WSNs). Three main contributions are identified in this thesis. First, we extend an optimization model that used the M/M/1 queueing model on a simple network consisting of a single source-destination pair by using the M/M/s queueing model on a general network consisting of several source-destination pairs. The model solves a multi-path routing problem by defining a Hamiltonian as a function of delay incurred and subjecting this Hamiltonian to Pontryagin's cost minimization to achieve efficient diffusion of traffic over the available parallel paths. Second, we revisit the problem of cost-based optimization in a multi-path setting by using a Game theoretical framework to propose and evaluate the performance of competitive and cooperative multi-path routing schemes and the impact of the routing metric (cost) on the difference between these two schemes. Finally, building upon a previously proposed optimization benchmark, we propose an Energy constrained QoS routing scheme for Wireless Sensor Networks and show through simulation that our scheme outperforms the benchmark scheme.

In recent years, many machine learning algorithms have been developed and widely applied in various applications. However, most of them have considered the data distributions of the training and test datasets to be similar. This thesis concerns on the decrease of generalization ability in a test dataset when the data distribution is different from that of the training dataset. As labels may be unavailable in the test dataset in practical applications, we follow the effective approach of unsupervised domain adaptation and propose distribution alignment methods to improve the generalization ability of models learned from the training dataset in the test dataset. To solve the problem of joint distribution alignment without target labels, we propose a new criterion of domain-shared group sparsity that is an equivalent condition for equal conditional distribution. A domain-shared group-sparse dictionary learning model is built with the proposed criterion, and a cross-domain label propagation method is developed to learn a target-domain classifier using the domain-shared group-sparse representations and the target-specific information from the target data. Experimental results show that the proposed method achieves good performance on cross-domain face and object recognition. Moreover, most distribution alignment methods have not considered the difference in distribution structures, which results in insufficient alignment across domains. Therefore, a novel graph alignment method is proposed, which aligns both data representations and distribution structural information across the source and target domains. An adversarial network is developed for graph alignment by mapping both source and target data to a feature space where the data are distributed with unified structure criteria. Promising results have been obtained in the experiments on cross-dataset digit and object recognition. Problem of dataset bias also exists in human pose estimation across datasets with different image qualities. Thus, this thesis proposes to synthesize target body parts for cross-domain distribution alignment, to address the problem of cross-quality pose estimation. A translative dictionary is learned to associate the source and target domains, and a cross-quality adaptation model is developed to refine the source pose estimator using the synthesized target body parts. We perform cross-quality experiments on three datasets with different image quality using two state-of-the-art pose estimators, and compare the proposed method with five unsupervised domain adaptation methods. Our experimental results show that the proposed method outperforms not only the source pose estimators, but also other unsupervised domain adaptation methods.

Made available in DSpace on 2015-03-03T11:52:20Z (GMT). No. of bitstreams: 0 Previous issue date: 2014-10-09Bitstream added on 2015-03-03T12:07:29Z : No. of bitstreams: 1 000808145.pdf: 874653 bytes, checksum: 2965900aa1ceda6337f055e33f179a25 (MD5) / O trabalho tem como propósito principal a análise da estabilidade a pequenas perturbações em Sistemas Elétricos de Potência, que é representado pelo Modelo Sensibilidade de Potência. São analisados três sistemas teste, conhecidos na literatura como Sul-Brasileiro, simétrico de duas áreas e New England. A princípio os sistemas teste foram submetidos a um ponto de operação de modo a serem instáveis. É proposta a utilização dos Estabilizadores de Sistema de Potência e do dispositivo Thyristor Controlled Series Capacitor em conjunto com o controlador Power Oscillation Damping. Estes controladores possuem a função de inserir amortecimento adicional às oscilações de baixa frequência do Sistema Elétrico de Potência. Neste caso é necessário alocar e ajustar os parâmetros dos respectivos controladores de forma correta. As localizações dos controladores são determinadas pelos fatores de participação (Estabilizadores de Sistema de Potência) e a distância entre polo de interesse e zero da função de transferência de malha aberta do controlador Power Oscillation Damping (conjunto Thyristor Controlled Series Capacitor-Power Oscillation Damping). Para o ajuste de parâmetros dos controladores (Estabilizadores de Sistema de Potência e Power Oscillation Damping) são utilizados dois algoritmos de otimização baseados em enxame de partículas, sendo eles o Particle Swarm Optimization e de Bacterial Foraging Optimization orientado pelas formas de movimentação do Particle Swarm Optimization. A atuação dos controladores alocados e ajustados de acordo com os métodos propostos aumenta o amortecimento dos modos oscilatórios de baixa frequência dos sistemas teste. Este fato é verificado pela análise dos resultados obtidos / This work aims to analyze small signal stability in power systems represented by the power sensitivity model. Three test systems known in the literature, such as South Brazilian, Two Symmetrical Areas and New England are analyzed. At first, the systems were submitted to an operation point to become unstable. The Power System Stabilizers and the device Thyristor Controlled Series Capacitor with the Power Oscillation Damping controller are proposed to use. These controllers insert additional damping to the low frequencies oscillations. In this case, it is necessary to allocate and adjust the parameters of the controllers correctly. The controller position is determined by the participation factors (Power System Stabilizers) and the distance between the interest pole and the zero of the open-loop transfer function of the Power Oscillation Damping Controller (Thyristor Controlled Series Capacitor-Power Oscillation Damping).Two algorithms based on the particle swarm are used to adjust the controller parameters (Power System Stabilizer and the Power Oscillation Damping), that are Particle Swarm Optimization and Bacterial Foraging Optimization oriented by the movement forms of the Particle Swarm Optimization. The actuation of the controller both allocated and adjusted according to the proposedmethod improves the low frequency oscillation damping. The results obtained confirm this conclusion

No description available.

Orientador: Percival Bueno de Araujo / Co-orientador: Laurence Duarte Colvara / Banca: Mara Lúcia Martins Lopes / Banca: Francisco Damasceno Freitas / Banca: Walmir de Freitas Filho / Resumo: O trabalho tem como propósito principal a análise da estabilidade a pequenas perturbações em Sistemas Elétricos de Potência, que é representado pelo Modelo Sensibilidade de Potência. São analisados três sistemas teste, conhecidos na literatura como Sul-Brasileiro, simétrico de duas áreas e New England. A princípio os sistemas teste foram submetidos a um ponto de operação de modo a serem instáveis. É proposta a utilização dos Estabilizadores de Sistema de Potência e do dispositivo Thyristor Controlled Series Capacitor em conjunto com o controlador Power Oscillation Damping. Estes controladores possuem a função de inserir amortecimento adicional às oscilações de baixa frequência do Sistema Elétrico de Potência. Neste caso é necessário alocar e ajustar os parâmetros dos respectivos controladores de forma correta. As localizações dos controladores são determinadas pelos fatores de participação (Estabilizadores de Sistema de Potência) e a distância entre polo de interesse e zero da função de transferência de malha aberta do controlador Power Oscillation Damping (conjunto Thyristor Controlled Series Capacitor-Power Oscillation Damping). Para o ajuste de parâmetros dos controladores (Estabilizadores de Sistema de Potência e Power Oscillation Damping) são utilizados dois algoritmos de otimização baseados em enxame de partículas, sendo eles o Particle Swarm Optimization e de Bacterial Foraging Optimization orientado pelas formas de movimentação do Particle Swarm Optimization. A atuação dos controladores alocados e ajustados de acordo com os métodos propostos aumenta o amortecimento dos modos oscilatórios de baixa frequência dos sistemas teste. Este fato é verificado pela análise dos resultados obtidos / Abstract: This work aims to analyze small signal stability in power systems represented by the power sensitivity model. Three test systems known in the literature, such as South Brazilian, Two Symmetrical Areas and New England are analyzed. At first, the systems were submitted to an operation point to become unstable. The Power System Stabilizers and the device Thyristor Controlled Series Capacitor with the Power Oscillation Damping controller are proposed to use. These controllers insert additional damping to the low frequencies oscillations. In this case, it is necessary to allocate and adjust the parameters of the controllers correctly. The controller position is determined by the participation factors (Power System Stabilizers) and the distance between the interest pole and the zero of the open-loop transfer function of the Power Oscillation Damping Controller (Thyristor Controlled Series Capacitor-Power Oscillation Damping).Two algorithms based on the particle swarm are used to adjust the controller parameters (Power System Stabilizer and the Power Oscillation Damping), that are Particle Swarm Optimization and Bacterial Foraging Optimization oriented by the movement forms of the Particle Swarm Optimization. The actuation of the controller both allocated and adjusted according to the proposedmethod improves the low frequency oscillation damping. The results obtained confirm this conclusion / Doutor

Orientador: Fabrício Aparecido Breve / Banca: Moacir Antonelli Ponti / Banca: Daniel Carlos Guimarães Pedronette / Resumo: As pesquisas realizadas para o Sequenciamento de Genomas, Proteômica, Sistemas Biológicos, Diagnósticos Médicos, entre outros, geram uma grande quantidade de dados, fazendo necessário o apoio de soluções computacionais para a análise e interpretação desses dados. A utilização de técnicas de Aprendizado de Máquina, para a extração de conhecimentos úteis dessas grandes quantidades de dados, tem sido amplamente discutida entre pesquisadores da Biologia e da Computação. O processo para se rotular todos os dados gerados pelas pesquisas biológicas, assim como em outras áreas, é difícil, caro e/ou demorado. Assim, buscar maneiras de se atingir uma grande acurácia com poucos dados rotulados torna-se uma tarefa importante e desafiadora. Nesse sentido, o Aprendizado SemiSupervisionado mostra-se como uma opção importante uma vez que utiliza dados rotulados e não rotulados para o treinamento, sendo uma categoria intermediária entre o Aprendizado Supervisionado e o Não Supervisionado. Diversas abordagens para algoritmos de Aprendizado Semi-Supervisionado são encontradas na literatura. Dentre elas, destacam-se os métodos baseados em grafos, que representam os dados de entrada como nós de um grafo cuja estrutura é utilizada para propagar informações de rótulos dos nós rotulados para os demais nós. Destaca-se ainda que a abordagem baseada em grafos possui uma grande fundamentação matemática e computacional. Nesse contexto, este trabalho apresenta uma análise comparativa de alguns algoritmos semi-supervisionados, baseados em grafos, quando aplicados a dados biológicos relacionados aos campos de estudos da Proteômica e Transcriptômica. Adicionalmente, o trabalho propõe um novo dataset com dados reais oriundos de pesquisas biológicas com o transcriptoma de formigas da espécie Mycocepurus goeldii. Alguns experimentos realizados com os algoritmos semi-supervisionados são apresentados, levando em consideração sua... / Abstract: Research conducted for the sequencing of genomes, Proteomics, Systems Biology, Medical Diagnostics, among others, generate a lot of data, making it necessary the support of computing solutions for the analysis and interpretation of such data. The possibility of using machine learning techniques to extract useful knowledge of these large amounts of data has been widely discussed among researchers of Biology and Computer Science. The process of labeling all data generated by biological research, as well as in other areas, is difficult, costly and / or time consuming. Thus, searching ways to achieve a high accuracy with few labeled data is an important and challenging task. Accordingly, the Semi-Supervised Learning shows up as an important option since it uses both labeled and unlabeled data for training, being an intermediate category between the Supervised and Unsupervised Learning. Several approaches to semi-supervised learning algorithms are found in the literature. Among them, the highlights are the graph-based methods, which represent the input data as nodes in a graph, which structure is used to propagate label information from labeled nodes to the other nodes. It is also noteworthy that the graph-based approach has a great mathematical and computational validity. In this context, this paper presents a comparative analysis of some semi-supervised algorithms based on graphs, when applied to biological data analysis related to the field of proteomics and transcriptomics studies. In addition, the paper proposes a new dataset with actual data from biological research with the transcriptome of the Mycocepurus goeldii species of ants. Some experiments performed with semi-supervised algorithms are presented, considering its efficacy when compared with a few supervised methods / Mestre

Este trabalho apresenta um novo algoritmo de diagnóstico distribuído em nível de sistema, Distributed Network Reachability (DNR). O algoritmo permite que cada nodo de uma rede particionável de topologia arbitrária determine quais porções da rede estão alcançáveis e inalcançáveis. DNR é o primeiro algoritmo de diagnóstico distribuído que permite a ocorrência de eventos dinâmicos de falha e recuperação de nodos e enlaces, inclusive com partições e healings da rede. O estado diagnosticado de um nodo é ou sem-falha ou inatingível; o estado diagnosticado de um enlace é ou sem-falha ou não-respondendo ou inatingível. O algoritmo consiste de três fases: teste, disseminação e cálculo de alcançabilidade. Durante a fase de testes cada enlace é testado por um de seus nodos adjacentes em intervalos de teste alternados. Após a detecção de um novo evento, o testador inicia a fase de disseminação, na qual a nova informação de diagnóstico é transmitida para os nodos alcançáveis. A cada vez que um novo evento é detectado ou informado, a terceira fase é executada, na qual um algoritmo de conectividade em grafos é empregado para calcular a alcançabilidade da rede. O algoritmo DNR utiliza o número mínimo de testes por enlace por rodada de testes e tem a menor latência possível de diagnóstico, assegurada pela disseminação paralela de eventos. A correção do algoritmo é provada formalmente. Uma prova de correção no arcabouço bounded correctness também foi elaborada, incluindo latência delimitada de diagnóstico, latência delimitada de inicialização e acuidade. Um simulador do algoritmo foi implementado. Experimentos foram executados em diversas topologias incluindo grafos aleatórios (k-vertex connected e Power-Law) bem como grafos regulares (meshes e hipercubos). Extensivos resultados de simulação de eventos dinâmicos de falha e recuperação em nodos e enlaces são apresentados. / This thesis introduces the new Distributed Network Reachability (DNR) algorithm, a distributed system-level diagnosis algorithm that allows every node of a partitionable general topology network to determine which portions of the network are reachable and unreachable. DNR is the first distributed diagnosis algorithm that works in the presence of network partitions and healings caused by dynamic fault and repair events. A node is diagnosed as either working or unreachable and a link is diagnosed either as working or unresponsive or unreachable. The algorithm is formally specified and consists of three phases: test, dissemination, and reachability computation. During the testing phase each link is tested by one of the adjacent nodes at alternating testing intervals. Upon the detection of a new event, the tester starts the dissemination phase, in which the new diagnostic information is received by every reachable node in the network. New events can occur before the dissemination completes. After a new event is detected or informed, a working node runs the third phase, in which a graph connectivity algorithm is employed to compute the network reachability. The algorithm employs the optimal number of tests per link per testing interval and the best possible diagnosis latency, assured by the parallel dissemination of event information. The correctness of the algorithm is proved, including the bounded diagnostic latency, bounded start-up and accuracy. Experimental results obtained from simulation are presented. Simulated topologies include random graphs (k-vertex connected and Power-Law) as well as regular graphs (meshes and hypercubes). Extensive simulation results of dynamic fault and repair events on nodes and links are presented.

Este trabalho apresenta um novo algoritmo de diagnóstico distribuído em nível de sistema, Distributed Network Reachability (DNR). O algoritmo permite que cada nodo de uma rede particionável de topologia arbitrária determine quais porções da rede estão alcançáveis e inalcançáveis. DNR é o primeiro algoritmo de diagnóstico distribuído que permite a ocorrência de eventos dinâmicos de falha e recuperação de nodos e enlaces, inclusive com partições e healings da rede. O estado diagnosticado de um nodo é ou sem-falha ou inatingível; o estado diagnosticado de um enlace é ou sem-falha ou não-respondendo ou inatingível. O algoritmo consiste de três fases: teste, disseminação e cálculo de alcançabilidade. Durante a fase de testes cada enlace é testado por um de seus nodos adjacentes em intervalos de teste alternados. Após a detecção de um novo evento, o testador inicia a fase de disseminação, na qual a nova informação de diagnóstico é transmitida para os nodos alcançáveis. A cada vez que um novo evento é detectado ou informado, a terceira fase é executada, na qual um algoritmo de conectividade em grafos é empregado para calcular a alcançabilidade da rede. O algoritmo DNR utiliza o número mínimo de testes por enlace por rodada de testes e tem a menor latência possível de diagnóstico, assegurada pela disseminação paralela de eventos. A correção do algoritmo é provada formalmente. Uma prova de correção no arcabouço bounded correctness também foi elaborada, incluindo latência delimitada de diagnóstico, latência delimitada de inicialização e acuidade. Um simulador do algoritmo foi implementado. Experimentos foram executados em diversas topologias incluindo grafos aleatórios (k-vertex connected e Power-Law) bem como grafos regulares (meshes e hipercubos). Extensivos resultados de simulação de eventos dinâmicos de falha e recuperação em nodos e enlaces são apresentados. / This thesis introduces the new Distributed Network Reachability (DNR) algorithm, a distributed system-level diagnosis algorithm that allows every node of a partitionable general topology network to determine which portions of the network are reachable and unreachable. DNR is the first distributed diagnosis algorithm that works in the presence of network partitions and healings caused by dynamic fault and repair events. A node is diagnosed as either working or unreachable and a link is diagnosed either as working or unresponsive or unreachable. The algorithm is formally specified and consists of three phases: test, dissemination, and reachability computation. During the testing phase each link is tested by one of the adjacent nodes at alternating testing intervals. Upon the detection of a new event, the tester starts the dissemination phase, in which the new diagnostic information is received by every reachable node in the network. New events can occur before the dissemination completes. After a new event is detected or informed, a working node runs the third phase, in which a graph connectivity algorithm is employed to compute the network reachability. The algorithm employs the optimal number of tests per link per testing interval and the best possible diagnosis latency, assured by the parallel dissemination of event information. The correctness of the algorithm is proved, including the bounded diagnostic latency, bounded start-up and accuracy. Experimental results obtained from simulation are presented. Simulated topologies include random graphs (k-vertex connected and Power-Law) as well as regular graphs (meshes and hypercubes). Extensive simulation results of dynamic fault and repair events on nodes and links are presented.