A suposta ubiquidade de sistemas decomponíveis foi interpretada por Holland (1975) como o principal motivo para o desempenho dos algoritmos genéticos (Genetic Algorithms (GAs)). A hipótese de Building Blocks (BBs) sugere que algoritmos genéticos mais eficientes poderiam ser implementados, contudo, apenas anos depois essas ideias puderam ser avaliadas experimentalmente no contexto de algoritmos de estimação de distribuição (Estimation of Distribution Algorithms (EDAs)). EDAs utilizam modelos probabilísticos, estimados a partir da população, para inferir características do espaço de busca que poderiam ser utilizadas para implementar operadores de reprodução mais eficazes. Tanto em problemas mono- quanto multi-objetivo, EDAs emergiram sob a premissa de que a eficácia dos operadores de reprodução seria proporcional à representatividade dos modelos probabilísticos utilizados. No entanto, estudos recentes tem demonstrado que a dificuldade em se construir modelos confiáveis pode tornar essa premissa inviável. Ou seja, para certos problemas de otimização os modelos probabilísticos utilizados seriam, em geral, de baixa qualidade e, portanto, não produziriam operadores eficazes. Esta tese trata das limitações encontradas na construção de modelos probabilísticos (linkage learning) sob a perspectiva da multimodalidade dos problemas em questão. A análise teórica considerou problemas aditivamente separáveis, enquanto a generalização das conclusões foi investigada em instâncias do modelo NK-landscapes e do problema da mochila multidimensional (Multidimensional Knapsack Problem (MKP)). Os resultados indicaram que a acurácia dos modelos probabilísticos é se relaciona inversamente ao grau de multimodalidade da função objetivo e que, em casos de extrema multimodalidade a construção de modelos probabilísticos confiáveis pode ser tornar infactível. Este resultado poderia inviabilizar o uso de EDAs no contexto multiobjetivo, devido a intrínseca multimodalidade de tais problemas. No entanto, observou-se que apesar da ausência de estatísticas confiáveis sobre cada uma das funções objetivo, a correlação entre elas se torna estatisticamente observável e útil aos operadores de reprodução na manutenção da diversidade e controle convergência da população. / The supposed ubiquity of nearly-decomposable systems was interpreted by Holland (1975) as the rationale for the performance of Genetic Algorithms (GAs), the Building Block (BB) hypothesis. His seminal studies suggest more efficient GAs as viable, but only later on his ideas have become practically tangible in the context of Estimation of Distribution Algorithms (EDAs). EDAs employ probabilistic modeling so as to infer properties of the search space (BBs) that could be useful for the effectiveness of reproduction operators. In both, single- and multi-objective contexts, EDAs have emerged on the assumption there is a correlation between how much information a model can conceive and how effective reproduction operators can be. However, more recent results suggest the difficulties in producing accurate linkage models can prevent such a relation to be true. In other words, for some optimization problems linkage learning might not be able to produce accurate linkage models, hence EDAs would not outperform GAs. This thesis addresses the limits of linkage learning in the context of single- and bi-objective problems, regarding the influence of multimodality on the accuracy of the linkage models and the efficiency of EDAs. A theoretical analysis was performed in terms of additively separable functions and general conclusions are assessed through experimentation with instances of the NK-model and the Multidimensional Knapsack Problem (MKP). The results indicated that the accuracy of the linkage models tends to decrease as a result of increasing multimodality, which weakens pairwise dependencies and might lead to pairwise independence in extreme cases. Since most EDAs rely on bivariate statistics to estimate multivariate distributions, their applicability is limited to optimization problems within a certain range of multimodality. In multi-objective problems, on the other hand, some EDAs have shown better performance than GAs, which seemed as a contradiction since multi-objective problems are inherently multimodal. Our results suggest that in such cases the correlation among the objective functions becomes statistically evident, as a consequence, linkage learning models such correlation instead of problems substructures, which is useful to obtain a better exploration of extreme regions of the objective space.

A suposta ubiquidade de sistemas decomponíveis foi interpretada por Holland (1975) como o principal motivo para o desempenho dos algoritmos genéticos (Genetic Algorithms (GAs)). A hipótese de Building Blocks (BBs) sugere que algoritmos genéticos mais eficientes poderiam ser implementados, contudo, apenas anos depois essas ideias puderam ser avaliadas experimentalmente no contexto de algoritmos de estimação de distribuição (Estimation of Distribution Algorithms (EDAs)). EDAs utilizam modelos probabilísticos, estimados a partir da população, para inferir características do espaço de busca que poderiam ser utilizadas para implementar operadores de reprodução mais eficazes. Tanto em problemas mono- quanto multi-objetivo, EDAs emergiram sob a premissa de que a eficácia dos operadores de reprodução seria proporcional à representatividade dos modelos probabilísticos utilizados. No entanto, estudos recentes tem demonstrado que a dificuldade em se construir modelos confiáveis pode tornar essa premissa inviável. Ou seja, para certos problemas de otimização os modelos probabilísticos utilizados seriam, em geral, de baixa qualidade e, portanto, não produziriam operadores eficazes. Esta tese trata das limitações encontradas na construção de modelos probabilísticos (linkage learning) sob a perspectiva da multimodalidade dos problemas em questão. A análise teórica considerou problemas aditivamente separáveis, enquanto a generalização das conclusões foi investigada em instâncias do modelo NK-landscapes e do problema da mochila multidimensional (Multidimensional Knapsack Problem (MKP)). Os resultados indicaram que a acurácia dos modelos probabilísticos é se relaciona inversamente ao grau de multimodalidade da função objetivo e que, em casos de extrema multimodalidade a construção de modelos probabilísticos confiáveis pode ser tornar infactível. Este resultado poderia inviabilizar o uso de EDAs no contexto multiobjetivo, devido a intrínseca multimodalidade de tais problemas. No entanto, observou-se que apesar da ausência de estatísticas confiáveis sobre cada uma das funções objetivo, a correlação entre elas se torna estatisticamente observável e útil aos operadores de reprodução na manutenção da diversidade e controle convergência da população. / The supposed ubiquity of nearly-decomposable systems was interpreted by Holland (1975) as the rationale for the performance of Genetic Algorithms (GAs), the Building Block (BB) hypothesis. His seminal studies suggest more efficient GAs as viable, but only later on his ideas have become practically tangible in the context of Estimation of Distribution Algorithms (EDAs). EDAs employ probabilistic modeling so as to infer properties of the search space (BBs) that could be useful for the effectiveness of reproduction operators. In both, single- and multi-objective contexts, EDAs have emerged on the assumption there is a correlation between how much information a model can conceive and how effective reproduction operators can be. However, more recent results suggest the difficulties in producing accurate linkage models can prevent such a relation to be true. In other words, for some optimization problems linkage learning might not be able to produce accurate linkage models, hence EDAs would not outperform GAs. This thesis addresses the limits of linkage learning in the context of single- and bi-objective problems, regarding the influence of multimodality on the accuracy of the linkage models and the efficiency of EDAs. A theoretical analysis was performed in terms of additively separable functions and general conclusions are assessed through experimentation with instances of the NK-model and the Multidimensional Knapsack Problem (MKP). The results indicated that the accuracy of the linkage models tends to decrease as a result of increasing multimodality, which weakens pairwise dependencies and might lead to pairwise independence in extreme cases. Since most EDAs rely on bivariate statistics to estimate multivariate distributions, their applicability is limited to optimization problems within a certain range of multimodality. In multi-objective problems, on the other hand, some EDAs have shown better performance than GAs, which seemed as a contradiction since multi-objective problems are inherently multimodal. Our results suggest that in such cases the correlation among the objective functions becomes statistically evident, as a consequence, linkage learning models such correlation instead of problems substructures, which is useful to obtain a better exploration of extreme regions of the objective space.

Didėjant interneto skvarbai ir visuomenės kompiuteriniam raštingumui, vis daugiau įmonių ir žmonių naudojasi elektroninėmis valdžios paslaugomis ir šių paslaugų teikimas tapo neatsiejama bet kurios valstybinės institucijos veiklos dalimi. EDAS sistemos įtaka Lietuvos verslui yra ypač svarbi, nes teikiamos socialinio draudimo paslaugos yra aktualios visoms įmonėms. E-valdžia įtakoja verslo sąlygas Lietuvoje ir tarptautinį valstybės konkurencingumą. Lietuvos Respublikos Vyriausybės nutarime „Dėl elektroninės valdžios koncepcijos patvirtinimo“ (2002 m. gruodžio 31 d. Nr. 2115) pažymima, kad „Lietuviškam verslui šis išbandymas gali būti labai sunkus, tačiau teigiami šio persiorientavimo rezultatai leis paprasčiau ir efektyviau dalyvauti pasaulinėse prekių ir paslaugų rinkose“. Vienas svarbiausių Lietuvos verslo sąlyčio taškų su valdžia yra Sodros veikla. Viena iš Sodros funkcijų yra duomenų apie draudėjus ir apdraustuosius apskaita vykdoma pagal Lietuvos verslo įmonių teikiamas ataskaitas. Kokybiškos elektroninės paslaugos leistų pagerinti Lietuvos verslo klimatą. Kyla klausimas kaip įvertinti EDAS sistemos atitikimą Lietuvos verslo lūkesčiams ir tuo pačiu kokios turėtų būti tolimesnės tobulinimo gairės. / Governments worldwide face with the challenge of transformation and the need to reinvent government systems in order to deliver efficient and cost effective services, information and knowledge through information and communication technologies. One of the most important aspects of e-government is how it moves businesses closer to the government. State Social Insurance Fund Board of the Republic of Lithuania presented the electronic services for Lithuanian business named EDAS. The object of this master thesis is e-government model application in the EDAS system. The aim of the work is to analyze theoretical e-government models and aspects, benchmark EDAS system according to evaluation criteria and to propose the guidance for improvement. The main tasks of this paper are to analyze the e-government models, application domains, aspects and criteria for benchmarking; to analyze scientific studies and international practice; to propose the benchmarking method which could evaluate EDAS system from a theoretical perspective; to measure Lithuanian business satisfaction with EDAS system using the survey and, according to its results, to propose the primary trends for improvements. The methodology for benchmarking was selected using logical e-government strengths and weaknesses analysis, American Customer Satisfaction Index criteria, Kearns e-government evaluation model, and European Commission e-Government Economics Project (eGEP). The aspects of these models were used for the... [to full text]

Vertical mixing is an important process that relates surface fluxes to concentrations of pollutants and other chemical species in the atmosphere. Errors in vertical mixing have been identified as a major source of uncertainties in various atmospheric modeling efforts including tracer transport, weather forecasting, and regional climate simulation. This thesis aims to quantify uncertainties in model-derived mixed layer heights (zi) over North America through direct comparisons between radiosonde observations and four models at different months of the year 2004 through the bulk Richardson number method. Results of this study suggest that considerable errors in zi exist throughout the region with the spatial and temporal variations of the errors differ significantly among the selected models. Over all, errors in zi were larger in global models than in the limited area mesoscale models, and the magnitude of the random error was two times larger than the bias. Notably, spatial regions of with extremely large positive biases correspond to those with especially large random errors. The biases and random errors, however, were not correlated linearly nor can be easily used to predict each other. Uncertainties in model-derived zi were attributed, through errors in the bulk Richardson number, to temperature and horizontal winds. Errors in both horizontal winds and temperatures were found contributing more or less the same to uncertainties in zi, with relative errors in both variables being the greatest in the lowest part of the troposphere. Lastly, independent observations from the cooperative profiler network suggest that data assimilation did not add qualitative advantages for the comparisons presented in this study. The mixed layer height uncertainties demonstrated in this study may provide a guide for selecting a model to simulate regional scale atmospheric transport and for interpreting flux estimation and inversions studies.

Vertical mixing is an important process that relates surface fluxes to concentrations of pollutants and other chemical species in the atmosphere. Errors in vertical mixing have been identified as a major source of uncertainties in various atmospheric modeling efforts including tracer transport, weather forecasting, and regional climate simulation. This thesis aims to quantify uncertainties in model-derived mixed layer heights (zi) over North America through direct comparisons between radiosonde observations and four models at different months of the year 2004 through the bulk Richardson number method. Results of this study suggest that considerable errors in zi exist throughout the region with the spatial and temporal variations of the errors differ significantly among the selected models. Over all, errors in zi were larger in global models than in the limited area mesoscale models, and the magnitude of the random error was two times larger than the bias. Notably, spatial regions of with extremely large positive biases correspond to those with especially large random errors. The biases and random errors, however, were not correlated linearly nor can be easily used to predict each other. Uncertainties in model-derived zi were attributed, through errors in the bulk Richardson number, to temperature and horizontal winds. Errors in both horizontal winds and temperatures were found contributing more or less the same to uncertainties in zi, with relative errors in both variables being the greatest in the lowest part of the troposphere. Lastly, independent observations from the cooperative profiler network suggest that data assimilation did not add qualitative advantages for the comparisons presented in this study. The mixed layer height uncertainties demonstrated in this study may provide a guide for selecting a model to simulate regional scale atmospheric transport and for interpreting flux estimation and inversions studies.