<p>It is rarely possible to use an optimal classifier. Often the classifier used for a specific problem is an approximation of the optimal classifier. Methods are presented for evaluating the performance of an approximation in the model class of Bayesian Networks. Specifically for the approximation of class conditional independence a bound for the performance is sharpened.</p><p>The class conditional independence approximation is connected to the minimum description length principle (MDL), which is connected to Jeffreys’ prior through commonly used assumptions. One algorithm for unsupervised classification is presented and compared against other unsupervised classifiers on three data sets.</p> / Report code: LiU-TEK-LIC 2006:11.

The present thesis introduces an iterative expert-based Bayesian approach for assessing greenhouse gas (GHG) emissions from the 2030 German new vehicle fleet and quantifying the impacts of their main drivers. A first set of expert interviews has been carried out in order to identify technologies which may help to lower car GHG emissions and to quantify their emission reduction potentials. Moreover, experts were asked for their probability assessments that the different technologies will be widely adopted, as well as for important prerequisites that could foster or hamper their adoption. Drawing on the results of these expert interviews, a Bayesian Belief Network has been built which explicitly models three vehicle types: Internal Combustion Engine Vehicles (which include mild and full Hybrid Electric Vehicles), Plug-In Hybrid Electric Vehicles, and Battery Electric Vehicles. The conditional dependencies of twelve central variables within the BBN - battery energy, fuel and electricity consumption, relative costs, and sales shares of the vehicle types - have been quantified by experts from German car manufacturers in a second series of interviews. For each of the seven second-round interviews, an expert's individually specified BBN results. The BBN have been run for different hypothetical 2030 scenarios which differ, e.g., in regard to battery development, regulation, and fuel and electricity GHG intensities. The present thesis delivers results both in regard to the subject of the investigation and in regard to its method. On the subject level, it has been found that the different experts expect 2030 German new car fleet emission to be at 50 to 65% of 2008 new fleet emissions under the baseline scenario. They can be further reduced to 40 to 50% of the emissions of the 2008 fleet though a combination of a higher share of renewables in the electricity mix, a larger share of biofuels in the fuel mix, and a stricter regulation of car CO$_2$ emissions in the European Union. Technically, 2030 German new car fleet GHG emissions can be reduced to a minimum of 18 to 44% of 2008 emissions, a development which can not be triggered by any combination of measures modeled in the BBN alone but needs further commitment. Out of a wealth of existing BBN, few have been specified by individual experts through elicitation, and to my knowledge, none of them has been employed for analyzing perspectives for the future. On the level of methods, this work shows that expert-based BBN are a valuable tool for making experts' expectations for the future explicit and amenable to the analysis of different hypothetical scenarios. BBN can also be employed for quantifying the impacts of main drivers. They have been demonstrated to be a valuable tool for iterative stakeholder-based science approaches. / Die vorliegende Arbeit verfolgt zwei Forschungsziele - ein inhaltliches und ein methodisches. Auf der inhaltlichen Ebene wurde die Entwicklung der CO2-Emissionen der deutschen Neuwagenflotte bis 2030 untersucht. Es wurden verschiedene technische Möglichkeiten daraufhin überprüft, inwieweit sie zur Emissionsminderung beitragen können, wie wahrscheinlich es ist, dass sie umgesetzt werden, und welche Voraussetzungen und Rahmenbedingungen bedeutenden Einfluß haben. Die methodische Innovation dieser Arbeit besteht darin, subjektive Einschätzungen von Experten mit einem Bayesianischen Netzwerk zu verknüpfen, um die Anwendung solcher Netzwerke auf Situationen von Unsicherheit im Knight'schen Sinne zu erweitern, hier am Beispiel der zukünftigen, heute nicht vorhersagbaren Entwicklung der CO2-Emissionen der deutschen Neuwagenflotte. Ein erster Schritt dieser Untersuchung bestand in der Erhebung und Auswertung der Einschätzungen von 15 Experten in Bezug auf die Möglichkeiten, die CO2-Emissionen neuer PKW in Deutschland bis 2020 zu senken. Erhoben wurden Aussagen über verfügbare Technologien, ihre Einsparpotenziale, ihre Umsetzungswahrscheinlichkeiten sowie wichtige Rahmenbedingungen. Ziel war es, wesentliche Variablen und deren Abhängigkeiten zu identifizieren, um eine Grundlage für die spätere Modellierung zu schaffen. Um die Untersuchung auf eine breite Basis zu stellen, wurden Experten von Autobauern und Zulieferern, Nichtregierungsorganisationen, Verbänden sowie solche aus Wissenschaft und Journalismus einbezogen. Aufbauend auf diesen Ergebnissen wurde in einem zweiten Schritt ein Bayesianisches Netzwerk entwickelt, mit dem die CO2-Emissionen der deutschen Neuwagenflotte im Jahr 2030 quantifiziert werden können. Außerdem sollten die Marktchancen verschiedener Fahrzeugtypen untersucht werden. Ein weiteres Ziel war es, den Einfluss verschiedener technologischer und regulatorischer Einflussfaktoren zu quantifizieren, die in der ersten Interviewrunde identifiziert worden waren. Gegenüber der ersten Interviewrunde wurde der zeitliche Rahmen der Untersuchung um 10 Jahre auf das Jahr 2030 erweitert. Das Bayesianische Netz erstreckt sich auf die zukünftigen Eigenschaften und Marktchancen von drei Fahrzeugtypen: Verbrennungsmotorische Fahrzeuge einschließlich aller Hybridvarianten bis hin zum Vollhybrid, Plug-In Hybride und Batterie-Elektrofahrzeuge. Das Netzwerk umfasst 46 miteinander verknüpfte Variablen. Für zwölf entscheidende Variablen wurden per Expertenbefragung bedingte Wahrscheinlichkeiten erhoben. Befragt wurden sieben Experten, fast alle hochrangige F&E- oder Umweltexperten bei deutschen Autobauern. Für jeden Experten entstand ein individuell quantifiziertes Netzwerk. Um mögliche Technologie- und Emissionspfade zu untersuchen, wurden verschiedene Szenarien definiert, die unterschiedliche Regulierungen, Batterie-Entwicklungspfade und CO2-Intensitäten von Treibstoffen und elektrischer Energie in Betracht ziehen. Im Basis-Szenario liegen die Erwartungswerte der CO2-Emissionen der deutschen Neuwagenflotte 2030 nach Einschätzung der Experten bei 50 bis 65% der Emissionen der Neuwagenflotte 2008. Kombiniert man einen gesteigerten Anteil erneuerbarer Energien im Strommix, einen größeren Anteil von Biotreibstoffen im Treibstoff-Mix und eine strengere CO2-Emissionsregulierung seitens der Europäischen Union, liegen die erwarteten Emissionen der Neuwagenflotte 2030 bei 40 bis 50% der Emissionen der Neuwagen 2008. Die Erwartungswerte der Neuflotten-Emissionen 2030 können in den verschiedenen BBN auf minimale Werte von 18 bis 44% der Emissionen der deutschen Neuwagenflotte von 2008 festgelegt werden. Es ist in den BBN durchaus möglich, aber in den betrachteten Szenarien unwahrscheinlich, die Emissionen so stark zu senken. Neben der Untersuchung von Fahrzeugtechnologien und CO2-Emissionen bis 2030 verfolgte die vorliegende Arbeit das Ziel, eine innovative Methode zu erproben. Es hat meines Wissens bisher kein Experten-basiertes BBN gegeben, das zukünftige Entwicklungen untersucht. Eine weitere Besonderheit des Ansatzes ist, dass für jeden Experten ein eigenes BBN quantifiziert wurde, so dass eine Schar von Netzwerken entstanden ist, deren Aussagen verglichen werden können. Dadurch war es möglich, die Bandbreite der Erwartungen verschiedener Experten abzuleiten und festzustellen, wo Erwartungen relativ konsistent sind, und wo sie weit auseinander liegen. Daraus lassen sich wertvolle Schlüsse ableiten, an welchen Punkten weitere Forschung besonders erfolgversprechend ist. BBN haben sich damit als nützliches Werkzeug für Stakeholder-basierte iterative Forschungsprozesse erwiesen. Insgesamt hat es sich als fruchtbarer neuer Ansatz erwiesen, Experten-basierte Bayesianische Netzwerke zur Untersuchung zukünftiger Entwicklungsmöglichkeiten heranzuziehen. Die Methode ermöglicht es auch, den Einfluß von Rahmenbedingungen zu bestimmen.

Adverse weather such as hurricanes can significantly affect the reliability of composite power systems. Predicting the impact of hurricanes can help utilities for better preparedness and make appropriate restoration arrangements. In this dissertation, the impact of hurricanes on the reliability of composite power systems is investigated. Firstly, the impact of adverse weather on the long-term reliability of composite power systems is investigated by using Markov cut-set method. The Algorithms for the implementation is developed. Here, two-state weather model is used. An algorithm for sequential simulation is also developed to achieve the same goal. The results obtained by using the two methods are compared. The comparison shows that the analytical method can obtain comparable results and meantime it can be faster than the simulation method. Secondly, the impact of hurricanes on the short-term reliability of composite power systems is investigated. A fuzzy inference system is used to assess the failure rate increment of system components. Here, different methods are used to build two types of fuzzy inference systems. Considering the fact that hurricanes usually last only a few days, short-term minimal cut-set method is proposed to compute the time-specific system and nodal reliability indices of composite power systems. The implementation demonstrates that the proposed methodology is effective and efficient and is flexible in its applications. Thirdly, the impact of hurricanes on the short-term reliability of composite power systems including common-cause failures is investigated. Here, two methods are proposed to archive this goal. One of them uses a Bayesian network to alleviate the dimensionality problem of conditional probability method. Another method extends minimal cut-set method to accommodate common-cause failures. The implementation results obtained by using the two methods are compared and their discrepancy is analyzed. Finally, the proposed methods in this dissertation are also applicable to other applications in power systems.

Stock market plays an important role in the modern capital market. As a result, the prediction of financial assets attracts people in different areas. Moreover, it is commonly accepted that stock price movement generally follows a major trend. As a result, forecasting the market trend becomes an important mission for a prediction method. Accordingly, we will predict the long term trend rather than the movement of near future or change in a trading day as the target of our predicting approach. Although there are various kinds of analyses for trend prediction, most of them use clear cuts or certain thresholds to classify the trends. Users (or investors) are not informed with the degrees of confidence associated with the recommendation or the trading signal. Therefore, in this research, we would like to study an approach that could offer the confidence of the trend analysis by providing the probabilities of each possible state given its historical data through Dynamic Bayesian Network. We will incorporate the well-known principles of Dow¡¦s Theory to better model the trend of stock movements. Through the results of our experiment, we may say that the financial performance of the proposed model is able to defeat the buy and hold trading strategy when the time scope covers the entire cycle of a trend. It also means that for the long term investors, our approach has high potential to win the excess return. At the same time, the trading frequency and correspondently trading costs can be reduced significantly.

In recent years, multi-agent traffic simulation has been widely used to accurately evaluate the performance of a road network considering individual and dynamic movements of vehicles under a virtual roadway environment. Given initial traffic demands and road conditions, the simulation is executed with multiple iterations and provides users with converged roadway conditions for the performance evaluation. For an accurate traffic simulation model, the driver's learning behavior is one of the major components to be concerned, as it affects road conditions (e.g., traffic flows) at each iteration as well as performance (e.g., accuracy and computational efficiency) of the traffic simulation. The goal of this study is to propose a realistic learning behavior model of drivers concerning their uncertain perception and interactions with other drivers. The proposed learning model is based on the Extended Belief-Desire-Intention (E-BDI) framework and two major decisions arising in the field of transportation (i.e., route planning and decision-making at an intersection). More specifically, the learning behavior is modeled via a dynamic evolution of a Bayesian network (BN) structure. The proposed dynamic learning approach considers three underlying assumptions: 1) the limited memory of a driver, 2) learning with incomplete observations on the road conditions, and 3) non-stationary road conditions. Thus, the dynamic learning approach allows driver agents to understand real-time road conditions and estimate future road conditions based on their past knowledge. In addition, interaction behaviors are also incorporated in the E-BDI framework to address influences of interactions on the driver's learning behavior. In this dissertation work, five major human interactions adopted from a social science literature are considered: 1) accommodation, 2) collaboration, 3) compromise, 4) avoidance, and 5) competition. The first three interaction types help to mimic information exchange behaviors between drivers (e.g., finding a route using a navigation system) while the last two interaction types are relevant with behaviors involving non-information exchange behaviors (e.g., finding a route based on a driver's own experiences). To calibrate the proposed learning behavior model and evaluate its performance in terms of inference accuracy and computational efficiency, drivers' decision data at intersections are collected via a human-in-the-loop experiment involving a driving simulator. Moreover, the proposed model is used to test and demonstrate the impact of five interactions on drivers' learning behavior under an en route planning scenario with real traffic data of Albany, New York, and Phoenix, Arizona. In this dissertation work, two major traffic simulation platforms, AnyLogic® and DynusT®, are used for the demonstration purposes. The experimental results reveal that the proposed model is effective in modeling realistic learning behaviors of drivers in conduction with interactions with other drivers.

Sharing components in a product family requires a trade-off between the individual products' performances and overall family costs. It is critical for a successful family to identify which components are similar, so that sharing does not compromise the individual products' performances. This research formulates two commonality identification approaches for use in product family design and investigates their applicability in a generic product family design methodology. Having a commonality identification approach reduces the combinatorial sharing problem and allows for more quality family alternatives to be considered. The first is based on the pattern recognition technique of fuzzy c-means clustering in component subspaces. If components from different products are similar enough to be grouped into the same cluster, then those components could possibly become the same platform. Fuzzy equivalence relations that show the binary relationship from one products' component to a different products' component can be extracted from the cluster membership functions. The second approach builds a Bayesian network representing the joint distribution of a design space exploration. Using this model, a series of inferences can be made based on product performance and component constraints. Finally the posterior design variable distributions can be processed using a similarity metric like the earth mover distance to identify which products' components are similar to another's.

京都大学 / 新制・課程博士 / 博士(薬学) / 甲第23844号 / 薬博第851号 / 新制||薬||242(附属図書館) / 京都大学大学院薬学研究科薬学専攻 / (主査)教授 山下 富義, 教授 石濱 泰, 教授 金子 周司 / 学位規則第4条第1項該当 / Doctor of Pharmaceutical Sciences / Kyoto University / DFAM

Thesis (MSc (Applied Mathematics))--University of Stellenbosch, 2010. / ENGLISH ABSTRACT: This thesis provides an introduction to the statistical modeling technique known as graphical models. Since graph theory and probability theory are the two legs of graphical models, these two topics are presented, and then combined to produce two examples of graphical models: Bayesian Networks and Markov Random Fields. Furthermore, the max-sum, sum-product and junction tree algorithms are discussed. The graphical modeling technique is then applied to the specific problem of finding coplanar points in stereo images, taken with an uncalibrated camera. Although it is discovered that graphical models might not be the best method, in terms of speed, to use for this appliation, it does illustrate how to apply this technique in a real-life problem. / AFRIKAANSE OPSOMMING: Hierdie tesis stel die leser voor aan die statistiese modelerings-tegniek genoemd grafiese modelle. Aangesien grafiek teorie en waarskynlikheidsleer die twee bene van grafiese modelle is, word hierdie areas aangespreek en dan gekombineer om twee voorbeelde van grafiese modelle te vind: Bayesian Netwerke en Markov Lukrake Liggaam. Die maks-som, som-produk en aansluitboom algoritmes word ook bestudeer. Nadat die teorie van grafiese modelle en hierdie drie algoritmes afgehandel is, word grafiese modelle dan toegepas op ’n spesifieke probleem— om punte op ’n gemeenskaplike vlak in stereo beelde te vind, wat met ’n ongekalibreerde kamera geneem is. Alhoewel gevind is dat grafiese modelle nie die optimale metode is om punte op ’n gemeenskaplike vlak te vind, in terme van spoed, word die gebruik van grafiese modelle wel ten toongestel met hierdie praktiese voorbeeld. / National Research Foundation (South Africa)

O raciocínio realizado em agentes BDI envolve essencialmente manipular três estruturas de dados representando suas crenças, desejos e intenções. Crenças de agentes BDI tradicionais não representam incerteza, e podem ser expressas como um conjunto fechado de literais ground. As restrições que indicam se um dado desejo é viável e pode ser adotado como uma intenção em agentes BDI tradicionais podem ser representadas como expressões lógicas sobre crenças. Dado que Redes Bayesianas permitem que representem-se informações com incerteza probabilisticamente, agentes BDI bayesianos as empregam para suportar incerteza em suas crenças. Em agentes BDI bayesianos, crenças representadas em Redes Bayesianas referem-se a estados de variáveis de eventos, possuindo probabilidades dinâmicas individuais que referem-se à incerteza. Os processos the constituem o raciocínio neste modelo de agente requerem mudanças a fim de acomodar esta diferença. Dentre estes processos, este trabalho concentra-se especificamente na seleção de desejos. Uma estratégia prévia para seleção de desejos é baseada em aplicar um limiar a probabilidades de crenças. Entretanto, tal abordagem impede que um agente selecione desejos condicionados em crenças cujas probabilidades estejam abaixo de um certo limiar, mesmo que tais desejos pudessem ser atingidos caso fossem selecionados. Para lidar com esta limitação, desenvolvemos três abordagens alternativas para seleção de desejos sob incerteza: Ranking Probabilístico, Loteria Viciada e Seleção Multidesejos Aleatória com Viés. Probability Ranking seleciona um desejo usando uma lista de desejos ordenados em ordem decrescente de probabilidade de pré-condição. Loteria Viciada seleciona um desejo usando um valor numérico aleatório e intervalos numéricos – associados a desejos – proporcionais às probabilidades de suas pré-condições. Seleção Multidesejos Aleatória com Viés seleciona múltiplos desejos usando valores numéricos aleatórios e considerando as probabilidades de suas pré-condições. Apresentamos exemplos, incluindo o agente Vigia, assim como experimentos envolvendo este, para mostrar como essas abordagens permitem que um agente às vezes selecione desejos cujas crenças pré-condições possuem probabilidades muito baixas. / The reasoning performed in BDI agents essentially involves manipulating three data structures representing their beliefs, desires and intentions. Traditional BDI agents’ beliefs do not represent uncertainty, and may be expressed as a closed set of ground literals. The constraints that indicate whether a given desire is viable and passive to be adopted as an intention in traditional BDI agents may be represented as logical expressions over beliefs. Given that Bayesian Networks allow one to represent uncertain information probabilistically, Bayesian BDI agents employ Bayesian Networks to support uncertainty in their beliefs. In Bayesian BDI agents, beliefs represented in Bayesian Networks refer to states of event variables, holding individual dynamic probabilities that account for the uncertainty. The processes that constitute reasoning in this agent model require changes in order to accomodate this difference. Among these processes, this work is specifically concerned with desire selection. A previous strategy for desire selection is based on applying a threshold on belief probabilities. However, such an approach precludes an agent from selecting desires conditioned on beliefs with probabilities below a certain threshold, even if those desires could be achieved if they were selected. To address this limitation, we develop three alternative approaches to desire selection under uncertainty: Probability Ranking, Biased Lottery and Multi-Desire Biased Random Selection. Probability Ranking selects a desire using a list of desires sorted in decreasing order of precondition probability. Biased Lottery selects a desire using one random numeric value and desire-associated numeric intervals proportional to the probabilities of the desires’ preconditions. Multi-Desire Biased Random Selection selects multiple desires using random numeric values and considering the probabilities of their preconditions. We present examples, including theWatchman agent, as well as experiments involving the latter, to show how these approaches allow an agent to sometimes select desires whose belief preconditions have very low probabilities.

O raciocínio realizado em agentes BDI envolve essencialmente manipular três estruturas de dados representando suas crenças, desejos e intenções. Crenças de agentes BDI tradicionais não representam incerteza, e podem ser expressas como um conjunto fechado de literais ground. As restrições que indicam se um dado desejo é viável e pode ser adotado como uma intenção em agentes BDI tradicionais podem ser representadas como expressões lógicas sobre crenças. Dado que Redes Bayesianas permitem que representem-se informações com incerteza probabilisticamente, agentes BDI bayesianos as empregam para suportar incerteza em suas crenças. Em agentes BDI bayesianos, crenças representadas em Redes Bayesianas referem-se a estados de variáveis de eventos, possuindo probabilidades dinâmicas individuais que referem-se à incerteza. Os processos the constituem o raciocínio neste modelo de agente requerem mudanças a fim de acomodar esta diferença. Dentre estes processos, este trabalho concentra-se especificamente na seleção de desejos. Uma estratégia prévia para seleção de desejos é baseada em aplicar um limiar a probabilidades de crenças. Entretanto, tal abordagem impede que um agente selecione desejos condicionados em crenças cujas probabilidades estejam abaixo de um certo limiar, mesmo que tais desejos pudessem ser atingidos caso fossem selecionados. Para lidar com esta limitação, desenvolvemos três abordagens alternativas para seleção de desejos sob incerteza: Ranking Probabilístico, Loteria Viciada e Seleção Multidesejos Aleatória com Viés. Probability Ranking seleciona um desejo usando uma lista de desejos ordenados em ordem decrescente de probabilidade de pré-condição. Loteria Viciada seleciona um desejo usando um valor numérico aleatório e intervalos numéricos – associados a desejos – proporcionais às probabilidades de suas pré-condições. Seleção Multidesejos Aleatória com Viés seleciona múltiplos desejos usando valores numéricos aleatórios e considerando as probabilidades de suas pré-condições. Apresentamos exemplos, incluindo o agente Vigia, assim como experimentos envolvendo este, para mostrar como essas abordagens permitem que um agente às vezes selecione desejos cujas crenças pré-condições possuem probabilidades muito baixas. / The reasoning performed in BDI agents essentially involves manipulating three data structures representing their beliefs, desires and intentions. Traditional BDI agents’ beliefs do not represent uncertainty, and may be expressed as a closed set of ground literals. The constraints that indicate whether a given desire is viable and passive to be adopted as an intention in traditional BDI agents may be represented as logical expressions over beliefs. Given that Bayesian Networks allow one to represent uncertain information probabilistically, Bayesian BDI agents employ Bayesian Networks to support uncertainty in their beliefs. In Bayesian BDI agents, beliefs represented in Bayesian Networks refer to states of event variables, holding individual dynamic probabilities that account for the uncertainty. The processes that constitute reasoning in this agent model require changes in order to accomodate this difference. Among these processes, this work is specifically concerned with desire selection. A previous strategy for desire selection is based on applying a threshold on belief probabilities. However, such an approach precludes an agent from selecting desires conditioned on beliefs with probabilities below a certain threshold, even if those desires could be achieved if they were selected. To address this limitation, we develop three alternative approaches to desire selection under uncertainty: Probability Ranking, Biased Lottery and Multi-Desire Biased Random Selection. Probability Ranking selects a desire using a list of desires sorted in decreasing order of precondition probability. Biased Lottery selects a desire using one random numeric value and desire-associated numeric intervals proportional to the probabilities of the desires’ preconditions. Multi-Desire Biased Random Selection selects multiple desires using random numeric values and considering the probabilities of their preconditions. We present examples, including theWatchman agent, as well as experiments involving the latter, to show how these approaches allow an agent to sometimes select desires whose belief preconditions have very low probabilities.