No description available.

Classification tasks are an integral part of science, industry, medicine, and business; being such a pervasive technique, its smallest improvement is valuable. Artificial Neural Network (ANN) is one of the strongest techniques used in many disciplines for classification. The ANN technique suffers from drawbacks such as intransparency in spite of its high prediction power. In this dissertation, motivated by learning styles in human brains, ANN’s shortcomings are assuaged and its learning power is improved. Self-Organizing Map (SOM), an ANN variation which has strong unsupervised power, and Feedforward ANN, traditionally used for classification tasks, are hybridized to solidify their benefits and help remove their limitations. These benefits are in two directions: enhancing ANN’s learning power, and improving decision-making. First, the proposed method, named Self-Organizing Error-Driven (SOED) Artificial Neural Network (ANN), shows significant improvements in comparison with usual ANNs. We show SOED is a more accurate, more reliable, and more transparent technique through experimentation with five famous benchmark datasets. Second, the hybridization creates space for inclusion of decision-making goals at the level of ANN’s learning. This gives the classifier the opportunity to handle the inconclusiveness of the data smarter and in the direction of decision-making goals. Through three case studies, naming 1) churn decision analytics, 2) breast cancer diagnosis, and 3) quality control decision making through thermal monitoring of additive manufacturing processes, this novel and cost-sensitive aspect of SOED has been explored and lead to much quantified improvement in decision-making.

Uninterrupted power supply is an important requirement for electric ship since it has to confront frequent travel and hostilities. However, the occurrence of faults in the shipboard power systems interrupts the power service continuity and leads to the severe damage on the electrical equipments. Faults need to be quickly detected and isolated in order to restore the power supply and prevent the massive cascading outage effect on the electrical equipments. This thesis presents an Artificial Neural Network (ANN) based method for the fault classification and location in MVDC shipboard power systems using the transient information in the fault voltage and current waveforms. The proposed approach is applied to the cable of an equivalent MVDC system which is simulated using PSCAD. The proposed method is efficient in detecting the type and location of DC cable faults and is not influenced by changes in electrical parameters like fault resistance and load.

One of the crucial challenges for Data Center (DC) operation is inefficient thermal management which leads to excessive energy waste. The information technology (IT) equipment and cooling systems of a DC are major contributors to power consumption. Additionally, failure of a DC cooling system leads to higher operating temperatures, causing critical electronic devices, such as servers, to fail which leads to significant economic loss. Improvements can be made in two ways, through (1) better design of a DC architecture and (2) optimization of the system for better heat transfer from hot servers. Row-based cooling is a suitable DC configuration that reduces energy costs by improving airflow distribution. Here, the IT equipment is contained within an enclosure that includes a cooling unit which separates cold and back chambers to eliminate hot air recirculation and cold air bypass, both of which produce undesirable airflow distributions. Besides, due to scalability, ease of implementation, and operational cost, row-based systems have gained in popularity for DC computing applications. However, a general thermal model is required to predict spatiotemporal temperature changes inside the DC and properly apply appropriate strategies. As yet, only primitive tools have been developed that are time-consuming and provide unacceptable errors during extrapolative predictions. We address these deficiencies by developing a rapid, adaptive, and accurate hybrid model by combining a DDM and the thermofluid transport relations to predict temperatures in a DC. Our hybrid model has low interpolative prediction errors below 0.7 oC and extrapolative errors less than one half of black-box models. Additionally, by changing the studied DC configuration such as cooling unit fans and severs locations, there are a few zones with prediction error more than 2 oC. Existing methods for cooling unit fault detection and diagnosis (FDD) are designed to successfully overcome individually occurring faults but have difficulty handling simultaneous faults. We apply a gray-box model involves a case study to detect and diagnose cooling unit fan and pump failure in a row-based DC cooling system. Fast detection of anomalous behavior saves energy and reduces operational costs by initiating remedial actions. Cooling unit fans and pumps are relatively low-reliability components, where the failure of one or more components can cause the entire system to overheat. Therefore, appropriate energy-saving strategies depend largely on the accuracy and timeliness of temperature prediction models. We used our gray-box model to produce thermal maps of the DC airspace for single as well as simultaneous failure conditions, which are fed as inputs for two different data-driven classifiers, CNN and RNN, to rapidly predict multiple simultaneous failures. Our FDD strategy can detect and diagnose multiple faults with accuracy as high as 100% while requiring relatively few simultaneous fault training data samples. / Thesis / Candidate in Philosophy

No description available.

Computational creativity is a new and developing field of artificial intelligence concerned with computational systems that either autonomously produce original and functional products, or that augment the ability of humans to do so. As the role of computers in our daily lives is continuing to expand, the need for such systems is becoming increasingly important. We introduce and document the development of a new “creative” system, called DARCI (Digital ARtist Communicating Intention), that is designed to autonomously create novel artistic images that convey linguistic concepts to the viewer. Within the scope of this work, the system becomes capable of creating non-photorealistic renderings of existing image compositions so that they convey the semantics of given adjectives. Ultimately, we show that DARCI is capable of producing surprising artifacts that are competitive, in some ways, with those produced by human artists. As with the development of any “creative” system, we are faced with the challenges of incorporating the philosophies of creativity into the design of the system, assessing the system's creativity, overcoming technical shortcomings of extant modern algorithms, and justifying the system within its creative domain (in this case, visual art). In meeting these challenges with DARCI, we demonstrate three broad contributions of the system: 1) the contribution to the field of computational creativity in the form of an original system, new approaches to achieving autonomy in creative systems, and new practical assessment methods; 2) the contribution to the field of computer vision in the form of new image features for affective image annotation and a new dataset; and 3) the contribution to the domain of visual art in the form of mutually beneficial collaborations and participation in several art galleries and exhibits.

Dynamic pricing is a difficult problem for machine learning. The environment is noisy, dynamic and has a measurable cost associated with exploration that necessitates that learning be done in short-time horizons. These short-time horizons force the learning algorithms to make pricing decisions based on scarce data. In this work, various machine learning algorithms are compared in the context of dynamic pricing. These algorithms include the Kalman filter, artificial neural networks, particle swarm optimization and genetic algorithms. The majority of these algorithms have been modified to handle the pricing problem. The results show that these adaptations allow the learning algorithms to handle the noisy dynamic conditions and to learn quickly.

The following work presents a new set of general methods for improving neural network accuracy on classification tasks, grouped under the label of classification-based methods. The central theme of these approaches is to provide problem representations and error functions that more directly improve classification accuracy than conventional learning and error functions. The CB1 algorithm attempts to maximize classification accuracy by selectively backpropagating error only on misclassified training patterns. CB2 incorporates a sliding error threshold to the CB1 algorithm, interpolating between the behavior of CB1 and standard error backpropagation as training progresses in order to avoid prematurely saturated network weights. CB3 learns a confidence threshold for each combination of training pattern and output class. This models an error function based on the performance of the network as it trains in order to avoid local overfit and premature weight saturation. PL1 is a point-wise local binning algorithm used to calibrate a learning model to output more accurate posterior probabilities. This algorithm is used to improve the reliability of classification-based networks while retaining their higher degree of classification accuracy. These approaches are demonstrated to be robust to a variety of learning parameter settings and have better classification accuracy than standard approaches on a variety of applications, such as OCR and speech recognition.

User click modeling in web search is most commonly done through probabilistic graphical models. Due to the successful use of machine learning techniques in other fields of research, it is interesting to evaluate how machine learning can be applied to click modeling. In this thesis, modeling is done using recurrent neural networks trained on a distributed representation of the state of the art user browsing model (UBM). It is further evaluated how extending this representation with a set of latent variables that are easily derivable from click logs, can affect the model's prediction performance. Results show that a model using the original representation does not perform very well. However, the inclusion of simple variables can drastically increase the performance regarding the click prediction task. For which it manages to outperform the two chosen baseline models, which themselves are well performing already. It also leads to increased performance for the relevance prediction task, although the results are not as significant. It can be argued that the relevance prediction task is not a fair comparison to the baseline functions, due to them needing more significant amounts of data to learn the respective probabilities. However, it is favorable that the neural models manage to perform quite well using smaller amounts of data. It would be interesting to see how well such models would perform when trained on far greater data quantities than what was used in this project. Also tailoring the model for the use of LSTM, which supposedly could increase performance even more. Evaluating other representations than the one used would also be of interest, as this representation did not perform remarkably on its own. / Klickmodellering av användare i söksystem görs vanligtvis med hjälp av probabilistiska modeller. På grund av maskininlärningens framgångar inom andra områden är det intressant att undersöka hur dessa tekniker kan appliceras för klickmodellering. Detta examensarbete undersöker klickmodellering med hjälp av recurrent neural networks tränade på en distribuerad representation av en populär och välpresterande klickmodell benämnd user browsing model (UBM). Det undersöks vidare hur utökandet av denna representation med statistiska variabler som enkelt kan utvinnas från klickloggar, kan påverka denna modells prestanda. Resultaten visar att grundrepresentationen inte presterar särskilt bra. Däremot har användningen av simpla variabler visats medföra drastiska prestandaökningar när det kommer till att förutspå en användares klick. I detta syfte lyckas modellerna prestera bättre än de två baselinemodeller som valts, vilka redan är välpresterande för syftet. De har även lyckats förbättra modellernas förmåga att förutspå relevans, fastän skillnaderna inte är lika drastiska. Relevans utgör inte en lika jämn jämförelse gentemot baselinemodellerna, då dessa kräver mycket större datamängder för att nå verklig prestanda. Det är däremot fördelaktigt att de neurala modellerna når relativt god prestanda för datamängden som använts. Det vore intressant att undersöka hur dessa modeller skulle prestera när de tränas på mycket större datamängder än vad som använts i detta projekt. Även att skräddarsy modellerna för LSTM, vilket borde kunna öka prestandan ytterligare. Att evaluera andra representationer än den som användes i detta projekt är också av intresse, då den använda representationen inte presterade märkvärdigt i sin grundform.

This thesis is a comparative study where the question is whether a neural network approach can outperform the principal component analysis (PCA) approach for predicting changes of interest rate curves. Today PCA is the industry standard model for predicting interest rate curves. Specifically the goal is to better understand the correlation structure between Swedish and European swap rates. The disadvantage with the PCA approach is that only the information contained in the covariance matrix can be used and not for example whether or not the curve might behave different depending on the current state. In other words, some information that might be quite important to the curve dynamic is lost in the PCA approach. This raises the question whether the lost information is important for prediction accuracy or not. As previously been shown by Alexei Kondratyev in the paper "Learning Curve Dynamics with Artificial Neural Networks", the neural network approach is able to use more information in the data and therefore has potential to outperform the PCA approach. Our thesis shows that the neural network approach is able to achieve the same or higher accuracy than PCA when performing long term predictions. The results show that the neural network model has potential to replace the PCA model, however, it is a more time consuming model. Higher accuracy can probably be achieved if the network is more optimized. / Det här är en jämförande studie där syftet är att undersöka hurvida noggrannare prediktioner kan uppnås genom att använda sig av artificiella neurala nätverk (ANN) istället för principalkomponentanalys (PCA) för att förutspå swapräntekurvor. PCA är idag industristandard för att förutspå räntekurvor. Specifikt är målet att bättre kunna förstå korrelationsstrukturen mellan de Svenska swapräntorna och de Europiska swapräntorna. En nackdel med PCA är att den enda tillgängliga informationen sparas i kovariansmatrisen. Det kan till exempel vara fallet att kurvan beter sig väldigt annorlunda beroende på om de nuvarande räntenivåerna är höga eller låga. Eftersom att sådan information går förlorad i PCA-modellen ligger intresset i att undersöka hur mycket noggrannare prediktionerna kan bli om man tar tillvara på ännu mer av informationen i datan. Som Alexei Kondratyev visar i rapporten "Learning Curve Dynamics with Artificial Neural Networks", så har ANN-modellen potential att ersätta PCA-modellen för att förutspå räntekurvor. I denna studie framgår det att ANN-modellen uppnår samma eller bättre resultat jämfört med PCA-modellen vid längre prediktioner.