The design of survivable storage systems involves inherent trade-oJ among properties such as performance, security, and availability. A toolbox of simple and accurate models of these properties allows a designer to make informed decisions. This report focuses on availability modeling. We describe two ways of extending the classic model of availability with a single "correlation parameter" to accommodate correlated failures. We evaluate the efficacy of the models by comparing their results with real measurements. We also show the use of the models as design decision tools: we analyze the effects of availability and correlation on the ordering of data distribution schemes and we investigate the placement of related files.

The research presented here focuses on modeling machine-learning performance. The thesis introduces Seer, a system that generates empirical observations of classification-learning performance and then uses those observations to create statistical models. The models can be used to predict the number of training examples needed to achieve a desired level and the maximum accuracy possible given an unlimited number of training examples. Seer advances the state of the art with 1) models that embody the best constraints for classification learning and most useful parameters, 2) algorithms that efficiently find maximum-likelihood models, and 3) a demonstration on real-world data from three domains of a practicable application of such modeling. The first part of the thesis gives an overview of the requirements for a good maximum-likelihood model of classification-learning performance. Next, reasonable design choices for such models are explored. Selection among such models is a task of nonlinear programming, but by exploiting appropriate problem constraints, the task is reduced to a nonlinear regression task that can be solved with an efficient iterative algorithm. The latter part of the thesis describes almost 100 experiments in the domains of soybean disease, heart disease, and audiological problems. The tests show that Seer is excellent at characterizing learning-performance and that it seems to be as good as possible at predicting learning

: Modeling fault behaviors such as fault occurrences and active/benign durations is an essential step to the design and evaluation of fault-tolerant controller computers. We use a beta-binomial distribution to model fault occurrences both in the presence and in the absence of environmentally-induced (thus common-cause) faults. A multinomial distribution is used to model fault active durations. The proposed model is validated by testing it against the data generated by a simulation program that mimics a common-cause fault environment. The model is then applied to the determination of an optimal time-redundancy recovery method for EMI-induced failures in an N-modular redundant controller computer, demonstrating its utility and power. 1 Introduction It is commonplace to use digital computers for the control and monitoring of safety-critical real-time systems, such as avionics systems and nuclear reactors, where controller-computer failures may be life-threatening. Massive hardware redun...

Background: Studies have revealed large variations in average health status across social, economic, and other groups. No study exists on the distribution of the risk of ill-health across individuals, either within groups or across all people in a society, and as such a crucial piece of total health inequality has been overlooked. Some of the reason for this neglect has been that the risk of death, which forms the basis for most measures, is impossible to observe directly and difficult to estimate.

(Under the direction of GARY G. KOCH.) Many studies involve a data structure which has some degree of clustering. Examples include complex sample surveys where the clusters are selected by cluster sampling, repeated measurement studies where the outcome measures evaluated for a given subject under two or more experimental and/or observational conditions may be viewed as a cluster, and animal litter studies where the clustered structure involves observations that have been taken on one experimental unit (i.e., a litter) at a given time. All analyses of clustered data structures should be based on viewlng the cluster as the basic unit of study. Quantities of interest in such studies are the mean value (n) of an observed response and the correlation coefficient (p), which reflects the strength of clustering in the data. The focus of methodological work (either by direct, maximum likelihood, or weighted least squares methods) for the analysis of clustered data has been on the two-stage nested design. Moreover, the parameter of interest has been primarily n. Although p has been estimated as a byproduct of the analysis procedure, the cluster-specific estimators have been the cluster means, n.. Extensions of these efforts are made by developing methodology for analyzing clustered data which is applicable to random effects models ofthe nested and/or classification types. The proposed methodology for the two-stage nested design, in particular, focuses on two parameter estimators within each cluster i-the ith cluster mean n. and the joint

(Under the direction of Barry H. Margolin) We investigate certain inferential aspects of mixtures of multinomial distributions, both in nonparametric and parametric contexts. As a nonparametric mixture model we propose a k-population finite mixture of binomial distributions, which can be applied to the analysis of noniid data generated from a series of toxicological experiments. A necessary and sufficient identifiability condition for the k-population finite mixture of binomials is obtained. The maximum likelihood estimates (MLE's) of the k-population finite mixture of binomials is computed via the EM algorithm (Dempster, Laird and Rubin, 1977), and the asymptotic properties of the MLE's are discussed. The identifiability condition is equivalent to the positive definiteness of the information matrix for the parameters. The MLE's and their sampling distributions, together with the data mentioned above, provide an empirical check of the statistical procedures proposed by Margolin, Kaplan and Zeiger (1981).

In this paper a correction of SGoF multitesting method for dependent tests is introduced. The correction is based in the beta-binomial model, and therefore the new method is called Beta-Binomial SGoF (or BB-SGoF). Main properties of the new method are established, and its practical implementation is discussed. BB-SGoF is illustrated through the analysis of two different real data sets on gene/protein expression levels. The performance of the method is investigated through simulations too. One of the main conclusions of the paper is that SGoF strategy may have much power even in the presence of possible dependences among the tests. 1

For a Web download or e-book: Your use of this publication shall be governed by the terms established by the vendor at the time you acquire this publication. The scanning, uploading, and distribution of this book via the Internet or any other means without the permission of the publisher is illegal and punishable by law. Please purchase only authorized electronic editions and do not participate in or encourage electronic piracy of copyrighted materials. Your support of others ’ rights is appreciated. U.S. Government Restricted Rights Notice: Use, duplication, or disclosure of this software and related documentation by the

For a Web download or e-book: Your use of this publication shall be governed by the terms established by the vendor at the time you acquire this publication. The scanning, uploading, and distribution of this book via the Internet or any other means without the permission of the publisher is illegal and punishable by law. Please purchase only authorized electronic editions and do not participate in or encourage electronic piracy of copyrighted materials. Your support of others ’ rights is appreciated. U.S. Government Restricted Rights Notice: Use, duplication, or disclosure of this software and related documentation by the