Copyright (c) 2009 Purdue Libraries All rights reserved. http://docs.lib.purdue.edu/ecetr Recent documents in ECE Technical Reports en-us Wed, 07 Oct 2009 18:40:43 PDT 3600 http://docs.lib.purdue.edu/ecetr/391 http://docs.lib.purdue.edu/ecetr/391 Tue, 06 Oct 2009 13:33:24 PDT Irregular applications, which manipulate complex, pointer-based data structures, are a promising target for parallelization. Recent studies have shown that these programs exhibit a kind of parallelism called amorphous data-parallelism. Prior approaches to parallelizing these applications, such as thread-level speculation and transactional memory, often obscure parallelism because they do not distinguish between the concrete representation of a data structure and its semantic state; they conflate metadata and data.Exploiting the semantic commutativity of methods in complex data structures is a promising approach to exposing more parallelism. Prior work has shown that abstract locks can be used to capture a subset of commutativity properties, however, abstract locks cannot uncover the parallelism in some complex data structures, such as kd-trees and union-find structures. In this paper, we propose a more flexible implementation of commutativity properties, called gatekeepers, which capture more complex commutativity conditions and thus expose more parallelism.We provide a formal definition of semantic commutativity and define conditions under which abstract locking can be applied and those under which gatekeeping is necessary.We present a quantitative study demonstrating the benefits of abstract locking and gatekeeping in amorphous data-parallel programs. We also present an efficient implementation of gatekeeping, which we evaluate on a real-world application. Milind Kulkarni http://docs.lib.purdue.edu/ecetr/390 http://docs.lib.purdue.edu/ecetr/390 Tue, 15 Sep 2009 12:36:44 PDT Studies on drug design datasets are continuing to grow. These datasets are usually known as hard modeled, having a large number of features and a small number of samples. The most common problems in the drug design area are of regression type. Committee machines (ensembles) have become popular in machine learning because of their high performance. In this study, dynamics of ensembles on regression related drug design problems are investigated on a big dataset collection. The study tries to determine the most successful ensemble algorithm, the base algorithm-ensemble pair having the best / worst results, the best successful single algorithm, and the similarities of algorithms according to their performances. We also discuss whether ensembles always generate better results than single algorithms. M. Fatih Amasyali http://docs.lib.purdue.edu/ecetr/389 http://docs.lib.purdue.edu/ecetr/389 Wed, 09 Sep 2009 11:55:55 PDT The sensor scheduling problem tries to select one out of multiple available sensors at each time step to minimize a weighted sum of all the estimation errors over a certain time horizon. The problem can be solved by enumerating all the possible schedules. The complexity of such an enumeration approach grows exponentially fast as the horizon increases. In this report, by introducing some numerical relaxation parameter, we develop an efficient way to compute a suboptimal sensor schedule. It is shown that by choosing the relaxation parameter small enough, the performance of the obtained suboptimal schedule can be made arbitrarily close to the optimal one. Wei Zhang http://docs.lib.purdue.edu/ecetr/388 http://docs.lib.purdue.edu/ecetr/388 Wed, 09 Sep 2009 10:52:15 PDT This paper presents and validates methods to extend reuse distance analysis of application locality characteristics to shared-memory multicore platforms by accounting for invalidation-based cache-coherence and inter-core cache sharing. Existing reuse distance analysis methods track the number of distinct addresses referenced between reuses of the same address by a given thread, but do not model the effects of data references by other threads. This paper shows several methods to keep reuse stacks consistent so that they account for invalidations and cache sharing, either as references arise in a simulated execution or at synchronization points. These methods are evaluated against a Simics-based coherent cache simulator running several OpenMP and transaction-based benchmarks. The results show that adding multicore-awareness substantially improves the ability of reuse distance analysis to model cache behavior, reducing the error in miss ratio prediction (relative to cache simulation for a specific cache size) by an average of 69% for per-core caches and an average of 84% for shared caches. Derek L. Schuff http://docs.lib.purdue.edu/ecetr/387 http://docs.lib.purdue.edu/ecetr/387 Tue, 28 Jul 2009 11:52:36 PDT Motivation: Recognizing that many intrinsically disordered regions in proteins play key roles in vital functions and also in some diseases, identification of the disordered regions has became a demanding process for structure prediction and functional characterization of proteins. Therefore, many studies have been motivated on accurate prediction of disorder. Mostly, machine learning techniques have been used for dealing with the prediction problem of disorder due to the capability of extracting the complex relationships and correlations hidden in large data sets.Results: In this study, a novel method, named Border Vector Detection and Extended Adaptation (BVDEA) was developed for predicting disorder as an alternative accurate classifier. The classifier performs the predictions by using three types of structural features belonging to proteins. For attesting the performance of the method, three computational learning techniques and eleven specific tools were used for comparison. Training was executed based on the data by 5-fold cross validation. When compared with the three learning methods of GRNN, LVQ and BVDA, the proposed method gives the best accuracy on classification. The BVDEA also provides faster and more robust learning as compared to the others. The new method provides a significant contribution to predicting disorder and order regions of proteins. Irem Ersoz Kaya http://docs.lib.purdue.edu/ecetr/386 http://docs.lib.purdue.edu/ecetr/386 Tue, 28 Jul 2009 11:14:33 PDT In regression applications, there is no single algorithm which performs well with all data since the performance of an algorithm depends on the dataset used. In practice, different algorithms / approaches are tried, and the best one is selected in each application. It is meaningful to ask whether there is a different way instead of running such tedious experiments. In meta learning studies, one investigates clues for the performance of an algorithm over a dataset using several features of the dataset. In this context, it is important to estimate which dataset features (meta features) are most significant for the performance of the algorithm.In the literature, meta learning studies mostly specialize to classification problems. In this study, meta regression problems are comprehensively studied on 3 big dataset collections (totally 181 datasets). New and existent meta features (about 300) are used. The relationships between the datasets and the algorithms are investigated. Several relations are found between meta features and related performances. The created meta datasets are made available to interested researchers. M. Fatih Amasyali http://docs.lib.purdue.edu/ecetr/385 http://docs.lib.purdue.edu/ecetr/385 Tue, 28 Jul 2009 06:45:37 PDT Compressive Sensing is a recently developed technique that exploits the sparsity of naturally occurring signals and images to solve inverse problems when the number of samples is less than the size of the original signal. We apply this technique to solve underdetermined inverse problems that commonly occur in remote sensing, including superresolution, image fusion and deconvolution. We use l1-minimization to develop algorithms that perform as well as or better than conventional methods for these problems. Our algorithms use a library of samples from similar images or a model for the image to be reconstructed to express the image as a sparse linear combination. A set of feature vectors is generated from the library or basis and is used to find the sparsest linear combination that matches the data using l1-minimization. Atul Divekar http://docs.lib.purdue.edu/ecetr/382 http://docs.lib.purdue.edu/ecetr/382 Mon, 27 Apr 2009 08:23:59 PDT Classification of nonlinearly separable data by nonlinear support vector machines is often a difficult task especially due to the necessity of a choosing a convenient kernel type. In this study, we propose a new classification method called support vector selection and adaptation (SVSA) that is applicable to both linearly and nonlinearly separable data in terms of some reference vectors generated by processing of support vectors obtained from the linear SVM. The method consists of two steps called selection and adaptation. In these two steps, once the support vectors are obtained by a linear SVM, some of them are rejected and others are selected and adapted to become the reference vectors. Classification is next carried out by using the K Nearest Neighbor Method (KNN) with the reference vectors. In the first step, all support vectors are classified by KNN with respect to the training data excluding the support vectors. The misclassified support vectors are rejected, and the remaining support vectors are chosen as the reference vectors. In the second step, the reference vectors are adapted by moving them towards to or away from the decision boundaries by the Learning Vector Quantization (LVQ) method. At the end of the adaptation process, the reference vectors are finalized. During classification, the class of each input vector is detected with the minimum distance rule in which the distances are calculated from the input vector to all the reference vectors. The SVSA method was experimented with some synthetic and real data, and the experimental results showed that the SVSA is competitive with the traditional SVM. Gulsen Taskin Kaya http://docs.lib.purdue.edu/ecetr/384 http://docs.lib.purdue.edu/ecetr/384 Mon, 27 Apr 2009 08:06:49 PDT This paper studies the discrete-time switched LQR (DSLQR) problem based on a dynamic programming approach. One contribution of this paper is the analytical characterization of both the value function and the optimal hybridcontrol strategy of the DSLQR problem. Their connections to the Riccati equation and the Kalman gain of the classical LQR problem are also discussed. Several interesting properties of the value functions are derived. In particular, we show that under some mild conditions, the family of finite-horizon value functions of the DSLQR problem is homogeneous (of degree 2), uniformly bounded over the unit ball, and converges exponentially fast to the infinitehorizon value function. Based on these properties, efficient algorithms are proposed to solve the finite-horizon and infinite-horizon DSLQR problems. More importantly, we establish conditions under which the strategies generated by the algorithms are stabilizing and suboptimal. These conditions are derived explicitly in terms of subsystem matrices and are thus very easy to verify. The proposed algorithms and the analysis provide a systematic way of solving the DSLQR problem with guaranteed closed-loop stability and suboptimal performance. Simulation results indicate that the proposed algorithms can efficiently solve not only specific but also randomly generated DSLQR problems, making the NP-hard problems numerically tractable. Wei Zhang http://docs.lib.purdue.edu/ecetr/383 http://docs.lib.purdue.edu/ecetr/383 Mon, 27 Apr 2009 07:55:44 PDT Classification of nonlinearly separable data by nonlinear support vector machines is often a difficult task especially due to the necessity of a choosing a convenient kernel type. In this study, we propose a new classification method called support vector selection and adaptation (SVSA) that is applicable to both linearly and nonlinearly separable data in terms of some reference vectors generated by processing of support vectors obtained from the linear SVM. The method consists of two steps called selection and adaptation. In these two steps, once the support vectors are obtained by a linear SVM, some of them are rejected and others are selected and adapted to become the reference vectors. Classification is next carried out by using the K Nearest Neighbor Method (KNN) with the reference vectors. In the first step, all support vectors are classified by KNN with respect to the training data excluding the support vectors. The misclassified support vectors are rejected, and the remaining support vectors are chosen as the reference vectors. In the second step, the reference vectors are adapted by moving them towards to or away from the decision boundaries by the Learning Vector Quantization (LVQ) method. At the end of the adaptation process, the reference vectors are finalized. During classification, the class of each input vector is detected with the minimum distance rule in which the distances are calculated from the input vector to all the reference vectors. The SVSA method was experimented with some synthetic and real data, and the experimental results showed that the SVSA is competitive with the traditional SVM. Gulsen Taskin Kaya