[pt] Lidar com processos não estacionários requer adaptação rápida e, simultaneamente, evitar esquecimento catastrófico de um comportamento passado. Duas técnicas de treinamento em redes neurais que satisfazem este requerimento são propostas, uma no âmbito de aprendizado com supervisão e outra relacionada à classe de redes não supervisionadas. Um novo algoritmo de treinamento supervisionada em redes multi camadas para modelagem de sistemas num contexto não estacionário é proposto. O ajuste de pesos na rede é determinado pela solução de problema de compromisso entre manter a performance para os dados antigos de treinamento e se ajustar para um novo comportamento representado nos dados mais recentes. Esta abordagem tem por base a formalização do problema como a minimização do erro de saída da rede sobre os padrões entrada-saída passados, sujeita a restrição de codificação do novo padrão dentro da tolerância estabelecida. Técnicas de programação não linear são utilizadas para resolver o problema de otimização gerado e calcular o novo conjunto de pesos. Alguns experimentos numéricos que comparam a performance do algoritmo proposto a de uma rede backpropagation são oferecidos. Um modelo de redes Fuzzy ART modulares para formação de categorias com hierarquia é também proposto, de tal forma que cada módulo agrupa os protótipos das categorias desenvolvidas na camada anterior. Dessa forma, os níveis hierárquicos sucessivamente mais altos captam uma visualização mais geral dos padrões de entrada enquanto os níveis inferiores aprendem categorias mais especificas. Propriedades interessantes da rede Fuzzy são herdadas pelo modelo proposto. Resultados teóricos relacionados às propriedades desta nova abordagem são apresentados, bem como experimentos numéricos que comprovam e ilustram as mesmas. / [en] The main issue when dealing with non-stationary processes is related to the requirement of fast adaptation while simultaneously preventing catastrophic damage to previously learned behavior. In this thesis, two on-line learning techniques, one for supervised and the other for unsupervised artificial neural networks, are proposed. A new supervised procedure to continuously adjust weights in a multi layered perceptrons neural networks is proposed. This methodology is tailored to be used in time varying (or non-stationary) models, eliminating the necessity of retraining. The main objective is to keep the error related to the latest income data within a pre established tolerance, while maximizing the information incorporated up to that point. By choosing a balance parameter, the designer is able to decide on the relevance that should be attributed to the new data. Non-linear programming techniques are used in order to properly solve this trade off optimization problem and on-line calculate the new weight set. Numerical results for real and non- real data are presented, illustrating the potentiality and properties of the proposed approach. A modular Fuzzy ART model for hierarchical categorization of data is proposed in such a manner that each module groups the prototypes of the categories developed in the previous module or layer. In this way, the hierarchical levels of successively higher layers learn a more general pattern from the input data while the lower layers learn more specific categories. Interesting properties of the component Fuzzy ART network also apply to this new hierarchical network model like, fast and stable learning of arbitrary sequences of analogical or binary input patterns and the dynamic creation of categories for new input patterns presentation. Some theoretical results related to properties of this new approach for clustering applications are presented, as well as some illustrative numerical results.

Abstract This thesis presents methods for modelling conditional variance and uncertainty of prediction at a query point on the basis of industrial process data. The introductory part of the thesis provides an extensive background of the examined methods and a summary of the results. The results are presented in detail in the original papers. The application presented in the thesis is modelling of the mean and variance of the mechanical properties of steel plates. Both the mean and variance of the mechanical properties depend on many process variables. A method for predicting the probability of rejection in a quali?cation test is presented and implemented in a tool developed for the planning of strength margins. The developed tool has been successfully utilised in the planning of mechanical properties in a steel plate mill. The methods for modelling the dependence of conditional variance on input variables are reviewed and their suitability for large industrial data sets are examined. In a comparative study, neural network modelling of the mean and dispersion narrowly performed the best. A method is presented for evaluating the uncertainty of regression-type prediction at a query point on the basis of predicted conditional variance, model variance and the effect of uncertainty about explanatory variables at early process stages. A method for measuring the uncertainty of prediction on the basis of the density of the data around the query point is proposed. The proposed distance measure is utilised in comparing the generalisation ability of models. The generalisation properties of the most important regression learning methods are studied and the results indicate that local methods and quadratic regression have a poor interpolation capability compared with multi-layer perceptron and Gaussian kernel support vector regression. The possibility of adaptively modelling a time-varying conditional variance function is disclosed. Two methods for adaptive modelling of the variance function are proposed. The background of the developed adaptive variance modelling methods is presented.

Cloud Computing presents a novel way for businesses to procure their IT needs. Its elasticity and on-demand provisioning enables a shift from capital expenditures to operating expenses, giving businesses the technological agility they need to respond to an ever-changing marketplace. The rapid adoption of Cloud Computing, however, poses a unique challenge to Cloud providers—their already very large electricity bill and carbon footprint will get larger as they expand; managing both costs is therefore essential to their growth. This thesis squarely addresses the above challenge. Recognizing the presence of Cloud data centers in multiple locations and the differences in electricity price and emission intensity among these locations and over time, we develop an optimization framework that couples workload distribution with time-varying signals on electricity price and emission intensity for financial and environmental benefits. The framework is comprised of an optimization model, an aggregate cost function, and 6 scheduling heuristics. To evaluate cost savings, we run simulations with 5 data centers located across North America over a period of 81 days. We use historical data on electricity price, emission intensity, and workload collected from market operators and research data archives. We find that our framework can produce substantial cost savings, especially when workloads are distributed both geographically and temporally—up to 53.35% on electricity cost, or 29.13% on carbon cost, or 51.44% on electricity cost and 13.14% on carbon cost simultaneously.

The worldwide population of patients with chronic kidney disease (CKD) is growing, with estimated prevalence at 12-15% of adults. Of particular concern are those with late stage CKD, defined as an estimated glomerular filtration rate (eGFR)of less than 30 ml/min/1.73m2, as they are susceptible to the highest risk of adverse outcomes such as progression to end stage kidney disease (ESKD), cardiovascular disease and all-cause mortality (1, 2). As such, late stage CKD patients are often managed in specialized clinics with set clinical targets, standardized education and multi-disciplinary care(3). A key clinical target for therapeutic intervention and prevention of the progression of CKD is blood pressure (BP) reduction(4). Yet, multiple relevant questions remain regarding the strength and nature of association of BP with clinical outcomes in late stage CKD. As the risks of hypotension-related complications are high in late stage CKD, it remains unclear whether strict BP control delays CKD progression in a real world clinic population(5). Furthermore, it is unclear how to appropriately specify the nature of the longitudinal association between BP and clinical outcomes of ESKD and mortality. The overall objective of this thesis is to examine the longitudinal association of BP and adverse clinical outcomes in a cohort of 1203 patients (mean eGFR 17.8 ml/min/1.73m2; mean of 6.7 BP measures per patient) with late stage CKD. In our first paper we examined the association of repeat measures of BP with CKD progression, defined as a decline in eGFR. When modeling eGFR using longitudinal linear regression, we found that its over-time trajectory was non-linear and that this trajectory was modified by BP; thus, we found a significant time-dependant association between BP and eGFR. When modeling time to eGFR decline ≥ 30% using Cox proportional hazards regression with categorized BP specified as a time-dependent exposure, BP was significantly associated with risk of eGFR decline; in particular, extremes of low and high systolic blood pressure (SBP) and high diastolic blood pressure (DBP) significantly increased the risk of eGFR decline. In our second paper, we examined different methods of modelling longitudinal BP and its association with time to mortality and ESKD. We found that elevations in SBP and DBP, in particular, when expressed as current (most recent visit), lag (previous visit), and cumulative exposure were significantly associated with increased risk of ESKD while low SBP (current, lag and cumulative exposure) was significantly associated with increased risk of mortality. Baseline BP measures were not statistically significantly associated with any outcomes. In patients with more moderate ranges of SBP (121-140) or DBP (60-85) at baseline, a subsequent rise to >160 or > 85 respectively, was associated with an increased risk of ESKD. Thus, longitudinal BP measures in late-stage CKD are significantly associated with adverse outcomes and convey important information beyond baseline BP measures.

I detta examensarbete presenteras en ny metod för att göra schemaläggningsbeslut inom processindustrin och samtidigt ta hänsyn till det dynamiska beteendet hos processen. En modell av en luftsepareringsanläggning som producerar kvävgas och utnyttjar ett rörligt elpris användes för att exemplifiera denna metod. Modellen omfattade en kryogenisk destillationskolonn med en integrerad återloppskokare /kondensator, en multiströms värmeväxlare, en kompressor, två turbiner och en kondensator. Den innehöll 5079 ekvationer och 437 differentiella variabler. Dynamisk optimering användes för att approximera det dynamiska beteendet hos processen vid skiftningar mellan olika driftpunkter. Den registrerade data utnyttjades sedan för att identifiera en reducerad modell som fångade det transienta beteendet hos relevanta processvariabler. Den reducerade modellen bestod av 525 ekvationer och 67 differentiella variabler. Den identifierade modellen visade på god matchning mellan relevanta processvariabler i de simulerade övergångarna och den reducerade modellen. Den reducerade modellen användes för att optimera schemaläggningen av luftsepareringsprocessen så att elkostnaden över en tredagars period minimerades. De optimala resultaten visade på en minskning av kostnaden på 2.6 % jämfört med en konstant produktionstakt. Schemat implementerades och simulerades i den fullt dynamiska modellen över de första 24 timmarna för att jämföra relevanta processvariabler med den reducerade modellen. Resultaten visade på god matchning mellan de båda modellerna. Detta examensarbete visar att en exakt reducerade modell kan användas för att snabbt hitta ett optimalt schema över ett större processystem. Detta genom att kraftigt minska systemets storlek utan att offra noggrannhet av det dynamiska beteendet. / A novel framework for making plant scheduling decisions while considering the plant process dynamics is presented in this thesis. A model of an air separation unit built to supply nitrogen gas and subject to time-varying electricity prices was used to illustrate this framework. The model includes a cryogenic distillation column with an integrated reboiler/condenser, a multi-stream heat exchanger, a compressor, two turbines, and a liquefier. It consisted of 5079 equations and 437 differential variables. The dynamic behavior of the process during operating point transitions was determined using dynamic optimization. This data were used to establish a reduced order dynamic model of the system which captures the transient behavior of relevant process variables. The reduced order model consisted of 525 equations and 67 differential variables. The identified model showed a good fit between the relevant process variables in the simulated transitions and the reduced order model. The air separation unit process schedule was optimized using the reduced order model to minimize electricity cost over a three day time horizon. The optimal result showed a 2.6 % reduction in electricity cost compared to a flat production rate. The optimal schedule was implemented and simulated in the full dynamic model for the first 24 hours to compare the relevant process variables to the reduced model predictions. The result displayed good match between the reduced model and the full dynamic model. This thesis shows that an accurate reduced order dynamic model can be used for quickly finding the optimal schedule of large process systems. This by greatly reducing the size and complexity of the system without sacrificing accuracy of the dynamic behavior. Furthermore, it also shows the economic benefits of the integrating scheduling and control to count for the dynamic behavior of the system.

Linear time-invariant models are widely used in the control community. They often serve as approximations of nonlinear systems. For control purposes linear approximations are often good enough since feedback control systems are inherently robust to model errors. In this thesis some of the possibilities for linear time-varying modeling are studied. In the thesis it is shown that the balanced truncation procedure can be applied to reduce the order of linear time-varying systems. Many of the attractive properties of balanced truncation for time-invariant systems can be generalized into the time-varying framework. For example, it is shown that a truncated input-output stable system will be input-output stable, and computable simple worst-case error bounds are derived. The method is illustrated with model reduction of a nonlinear diesel exhaust catalyst model. It is also shown that linear time-periodic models can be used for analysis of systems with power converters. Power converters produce harmonics in the power grids and give frequency coupling that cannot be modeled with standard time-invariant linear models. With time-periodic models we can visualize the coupling and also use all the available tools for linear time-varying systems, such as balanced truncation. The method is illustrated on inverter locomotives. / QC 20120208

Travel demand management policies such as congestion charges encourage car-users to change among other things route, mode and departure time. Departure time may be especially affected by time-varying charges, since car-users can avoid high peak hour charges by travelling earlier or later, so called peak spreading effects. Conventional transport models do not include departure time choice as a response. For evaluation of time-varying congestion charges departure time choice is essential. In this thesis a transport model called SILVESTER is implemented for Stockholm. It includes departure time, mode and route choice. Morning trips, commuting as well as other trips, are modelled and time is discretized into fifteen-minute time periods. This way peak spreading effects can be analysed. The implementation is made around an existing route choice model called CONTRAM, for which a Stockholm network already exists. The CONTRAM network has been in use for a long time in Stockholm and an origin-destination matrix calibrated against local traffic counts and travel times guarantee local credibility. On the demand side, an earlier developed departure time and mode choice model of mixed logit type is used. It was estimated on CONTRAM travel times to be consistent with the route choice model. The behavioural response under time-varying congestion charges was estimated from a hypothetical study conducted in Stockholm. Paper I describes the implementation of SILVESTER. The paper shows model structure, how model run time was reduced and tests of convergence. As regards run time, a 75% cut down was achieved by reducing the number of origin-destination pairs while not changing travel time and distance distributions too much. In Paper II car-users underlying preferred departure times are derived using a method called reverse engineering. This method derives preferred departure times that reproduce as well as possible the observed travel pattern of the base year. Reverse engineering has previously only been used on small example road networks. Paper II shows that application of reverse engineering to a real-life road network is possible and gives reasonable results. / <p>QC 20170222</p> / Silvester

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第18701号 / 情博第551号 / 新制||情||97(附属図書館) / 31634 / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 阿久津 達也, 教授 山本 章博, 教授 岡部 寿男 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

No description available.

Understanding the effect of human-induced pollution on the environment is an important precursor to promoting public health and environmental stability. One aspect of understanding pollution is understanding pollution sources. Various methods have been used and developed to understand pollution sources and the amount of pollution those sources emit. Multivariate receptor modeling seeks to estimate pollution source profiles and pollution emissions from concentrations of pollutants such as particulate matter (PM) in the air. Previous approaches to multivariate receptor modeling make the following two key assumptions: (1) PM measurements are independent and (2) source profiles are constant through time. Notwithstanding these assumptions, the existence of temporal correlation among PM measurements and time-varying source profiles is commonly accepted. In this thesis an approach to multivariate receptor modeling is developed in which the temporal structure of PM measurements is accounted for by modeling source profiles as a time-dependent Dirichlet process. The Dirichlet process (DP) pollution model developed herein is evaluated using several simulated data sets. In the presence of time-varying source profiles, the DP model more accurately estimates source profiles and source contributions than other multivariate receptor model approaches. Additionally, when source profiles are constant through time, the DP model outperforms other pollution receptor models by more accurately estimating source profiles and source contributions.