Combining embedded systems and machine learning models is an exciting prospect. However, to fully target any embedded system, with the most stringent resource requirements, the models have to be designed with care not to overwhelm it. Decision tree ensembles are targeted in this thesis. A benchmark model is created with LightGBM, a popular framework for gradient boosted decision trees. This model is first transformed and regularized with RuleFit, a LASSO regression framework. Then it is further optimized with quantization and weight sharing, techniques used when compressing neural networks. The entire process is combined into a novel framework, called ESRule. The data used comes from the domain of frequency measurements in cellular networks. There is a clear use-case where embedded systems can use the produced resource optimized models. Compared with LightGBM, ESRule uses 72ˆ less internal memory on average, simultaneously increasing predictive performance. The models use 4 kilobytes on average. The serialized variant of ESRule uses 104ˆ less hard disk space than LightGBM. ESRule is also clearly faster at predicting a single sample.

Thesis (M.Sc. (Computer Science))--North-West University, Potchefstroom Campus, 2008.

If customers respond differently to a campaign, it is worthwhile to find those customers who respond most positively and direct the campaign towards them. This can be done by using so called incremental response analysis where respondents from a campaign are compared with respondents from a control group. Customers with the highest increased response from the campaign will be selected and thus may increase the company’s return. Incremental response analysis is applied to the mobile operator Tres historical data. The thesis intends to investigate which method that best explain the incremental response, namely to find those customers who give the highest incremental response of Tres customers, and what characteristics that are important.The analysis is based on various classification methods such as logistic regression, Lassoregression and decision trees. RMSE which is the root mean square error of the deviation between observed and predicted incremental response, is used to measure the incremental response prediction error. The classification methods are evaluated by Hosmer-Lemeshow test and AUC (Area Under the Curve). Bayesian logistic regression is also used to examine the uncertainty in the parameter estimates.The Lasso regression performs best compared to the decision tree, the ordinary logistic regression and the Bayesian logistic regression seen to the predicted incremental response. Variables that significantly affect the incremental response according to Lasso regression are age and how long the customer had their subscription.

In the last one hundred years, the mankind has invented a lot of convenient tools for pursuing beautiful and comfortable living environment. Computer is one of the most important inventions, and its operation ability is incomparable with the mankind. Because computer can deal with a large amount of data fast and accurately, people use this advantage to imitate human thinking. Artificial intelligence is developed extensively. Methods, such as all kinds of neural networks, data mining, fuzzy logic, etc., apply to each side fields (ex: fingerprint distinguishing, image compressing, antennal designing, etc.). We will probe into to prediction technology according to the decision tree and fuzzy clustering. The fuzzy decision tree proposed the classification method by using fuzzy clustering method, and then construct out the decision tree to predict for data. However, in the distance function, the impact of the target space was proportional inversely. This situation could make problems in some dataset. Besides, the output model of each leaf node represented by a constant restricts the representation capability about the data distribution in the node. We propose a more reasonable definition of the distance function by considering both input and target differences with weighting factor. We also extend the output model of each leaf node to a local linear model and estimate the model parameters with a recursive SVD-based least squares estimator. Experimental results have shown that our improved version produces higher recognition rates and smaller mean square errors for classification and regression problems, respectively.

Unscheduled aircraft maintenance causes a lot problems and costs for aircraft operators. This is due to the fact that aircraft cause significant costs if flights have to be delayed or canceled and because spares are not always available at any place and sometimes have to be shipped across the world. Reducing the number of unscheduled maintenance is thus a great costs factor for aircraft operators. This thesis describes three methods for aircraft health monitoring and prediction; one method for system monitoring, one method for forecasting of time series and one method that combines the two other methods for one complete monitoring and prediction process. Together the three methods allow the forecasting of possible failures. The two base methods use decision trees for decision making in the processes and genetic optimization to improve the performance of the decision trees and to reduce the need for human interaction. Decision trees have the advantage that the generated code can be fast and easily processed, they can be altered by human experts without much work and they are readable by humans. The human readability and modification of the results is especially important to include special knowledge and to remove errors, which the automated code generation produced.

Thesis (M.Sc. (Computer Science))--North-West University, Potchefstroom Campus, 2008.

Thesis (M.Sc. (Computer Science))--North-West University, Potchefstroom Campus, 2008.

Thesis (M.S.) -- Mississippi State University. Department of Computer Science and Engineering. / Title from title screen. Includes bibliographical references.

Cette Thèse pose la question de la décision collective sous incertitude possibiliste. On propose différents règles de décision collective qualitative et on montre que dans un contexte possibiliste, l'utilisation d'une fonction d'agrégation collective pessimiste égalitariste ne souffre pas du problème du Timing Effect. On étend ensuite les travaux de Dubois et Prade (1995, 1998) relatifs à l'axiomatisation des règles de décision qualitatives (l'utilité pessimiste) au cadre de décision collective et montre que si la décision collective comme les décisions individuelles satisfont les axiomes de Dubois et Prade ainsi que certains axiomes relatifs à la décision collective, particulièrement l'axiome de Pareto unanimité, alors l'agrégation collective égalitariste s'impose. Le tableau est ensuite complété par une axiomatisation d'un pendant optimiste de cette règle de décision collective. Le système axiomatique que nous avons développé peut être vu comme un pendant ordinal du théorème de Harsanyi (1955). Ce résultat á été démontré selon un formalisme qui et basé sur le modèle de de Von NeuMann and Morgenstern (1948) et permet de comparer des loteries possibilistes. Par ailleurs, on propose une première tentative pour la caractérisation des règles de décision collectives qualitatives selon le formalisme de Savage (1972) qui offre une représentation des décisions par des actes au lieu des loteries. De point de vue algorithmique, on considère l'optimisation des stratégies dans les arbres de décision possibilistes en utilisant les critères de décision caractérisés dans la première partie de ce travail. On offre une adaptation de l'algorithme de Programmation Dynamique pour les critères monotones et on propose un algorithme de Programmation Multi-dynamique et un algorithme de Branch and Bound pour les critères qui ne satisfont pas la monotonie. Finalement, on établit une comparaison empirique des différents algorithmes proposés. On mesure les CPU temps d'exécution qui augmentent linéairement en fonction de la taille de l'arbre mais restent abordable même pour des grands arbres. Ensuite, nous étudions le pourcentage d'exactitude de l'approximation des algorithmes exacts par Programmation Dynamique: Il apparaît que pour le critère U-max ante l'approximation de l'algorithme de Programmation Multi-dynamique n'est pas bonne. Mais, ceci n'est pas si dramatique puisque cet algorithme est polynomial (et efficace dans la pratique). Cependant, pour la règle U+min ante l'approximation par Programmation Dynamique est bonne et on peut dire qu'il devrait être possible d'éviter une énumération complète par Branch and Bound pour obtenir les stratégies optimales. / This Thesis raises the question of collective decision making under possibilistic uncertainty. We propose several collective qualitative decision rules and show that in the context of a possibilistic representation of uncertainty, the use of an egalitarian pessimistic collective utility function allows us to get rid of the Timing Effect. Making a step further, we prove that if both the agents' preferences and the collective ranking of the decisions satisfy Dubois and Prade's axioms (1995, 1998) and some additional axioms relative to collective choice, in particular Pareto unanimity, then the egalitarian collective aggregation is compulsory. The picture is then completed by the proposition and the characterization of an optimistic counterpart of this pessimistic decision rule. Our axiomatic system can be seen as an ordinal counterpart of Harsanyi's theorem (1955). We prove this result in a formalism that is based on Von NeuMann and Morgenstern framework (1948) and compares possibilisitc lotteries. Besides, we propose a first attempt to provide a characterization of collective qualitative decision rules in Savage's formalism; where decisions are represented by acts rather than by lotteries. From an algorithmic standpoint, we consider strategy optimization in possibilistic decision trees using the decision rules characterized in the first part of this work. So, we provide an adaptation of the Dynamic Programming algorithm for criteria that satisfy the property of monotonicity and propose a Multi-Dynamic programming and a Branch and Bound algorithm for those that are not monotonic. Finally, we provide an empirical comparison of the different algorithms proposed. We measure the execution CPU times that increases linearly according to the size of the tree and it remains affordable in average even for big trees. Then, we study the accuracy percentage of the approximation of the pertinent exact algorithms by Dynamic Programming: It appears that for U-max ante criterion the approximation of Multi-dynamic programming is not so good. Yet, this is not so dramatic since this algorithm is polynomial (and efficient in practice). However, for U+min ante decision rule the approximation by Dynamic Programming is good and we can say that it should be possible to avoid a full Branch and Bound enumeration to find optimal strategies.

No Brasil, a execução de mapeamento de solos em todo o território nacional é uma demanda permanente das instituições de pesquisa e por órgãos de planejamento, dado que é uma importante ferramenta para o planejamento da ocupação racional das terras. O Mapeamento Digital de Solo (MDS) surge como alternativa para aumentar a viabilidade de execução de levantamentos de solos, utilizando-se de informações relacionadas ao relevo para mapear os solos. Este estudo objetiva testar metodologias de MDS com extrapolação para área fisiografimente semelhante e reclassificar o mapa pedológico gerado por MDS para criar um mapa de aptidão agrícola das terras e compará-lo com o mapa interpretativo gerado a partir do mapa convencional. Tendo em vista a escassez de dados existentes na Encosta do Sudeste do Rio Grande do Sul, o trabalho foi realizado em Sentinela do Sul e Cerro Grande do Sul. O MDS usou como modelos preditores um modelo geral de árvore de decisão (AD), testando-se um modelo para toda área e também o uso conjunto de dois modelos de predição. Uma vez que o MDS mapeia normalmente classes e propriedades dos solos e que desconhece-se o uso de tal técnica para gerar mapas de aptidão agrícola das terras, parte-se da hipótese que estes mapas possam ser criados a partir da reclassificação do mapa de solos gerados por MDS. O uso de modelos conjuntos de AD gerou modelos com mais acertos e maior capacidade de reprodução do mapa convencional de solos. A extrapolação para o município de Cerro Grande do Sul se mostrou eficiente. Ao classificar a aptidão agrícola das terras, a concordância entre o mapa convencional e os mapas preditos foi maior do que a concordância entre os mapas de solos. / In Brazil, the implementation of soil mapping throughout the national territory is a constant demand of research institutions and planning organs, as it is an important tool for rational planning of land occupation. Digital Soil Mapping (DSM) is an alternative to increase the viability of the soil survey because plots the information based on the relief to draw the soil map. This study aims to test methodologies DSM applied to similar landscapes areas. It also aims to reclassify the pedological map generated by DSM to create a new land suitability classes map and compare it with the land suitability classes map generated from conventional maps. The study was conducted in South Sentinel and Cerro Grande do Sul considering the lack of data in that area. The MDS was generated using a global model of decision tree (DT) for the entire area and combined with the use of two predictive models. The use of DSM to land suitability classes map is unknown. Perhaps interpretive maps created from the reclassification of DSM can produce more accurate maps than the predictor model would generate of the pedological map. The use of set models of DT created models with greater hits and higher reproductive capacity of the conventional map. The extrapolation to Cerro Grande do Sul was efficient . The DSM was more efficient to classify land suitability classes than to classify pedological maps, but this system of land sutability needs adjustments to reflect the local reality.