Browsing by Subject "Neural networks (Computer science)"
Now showing 1 - 20 of 28
Results Per Page
Sort Options
Item A machine learning approach to automate classification of literature in a SAM research database(Texas Tech University, 2004-08) Morris, Elizabeth PIn the mid-eighties, researchers at the University of Miami confronted their problem of information overload while investigating information on worker performance. They required literature sources from various fields, such as engineering, business, and psychology, to name a few. To cope with their information overload, they devised a research methodology to partition information resources into category matrices in order to find pattems, frends, or voids. The approach was termed State-of-the-Art Matrix or SAM Analysis. SAM Analysis is a manual process, thus restricting the amount of information for conveying category decisions. During the first phase of the manual process, researchers construct models or categories that best describe the research area. In the next phase, articles from the information sources are read and assigned to the pre-defined categories based on the judgment of assessors. The manual approach presents major challenges to researchers who must deal with identifying and utilizing the information hidden in a large corpus of information. The approach is only practical for a small number of articles and categorization relies on the subjective judgment of assessors. A more scalable and flexible approach, therefore, is needed for categorizing information, such as by using machine leaming and data mining techniques to automate categorization of articles in large volumes of data. In this research, automation is approached through the use of a machine leaming technique known as a Leaming Classifier Systems (LCS). The LCS performs the data mining task of categorizing articles using the SAM approach by utilizing training and testing datasets extracted from SAM EndNote bibliographic databases related to a specific area of research. In order to evaluate the ability of the LCS to predict category membership, accuracy-based metrics borrowed from the field of medicine are applied. The metrics include sensitivity, specificity, positive predictive value, and negative predictive value. After training, the evaluation results indicate that the predictive ability of the LCS system is greater than 90%. The results are obtained during the second trial of a five trial experiment.Item A multibit cascaded sigma-delta modulator with DAC error cancellation techniques(Texas Tech University, 2004-05) Su, Chun-hsienNoise reduction techniques are developed for a multibit cascaded sigma-delta (ÓÄ) modulator used in the analog interface of a digital signal processing system to improve its performance by reducing the errors introduced by digital-to-analog converters (DACs). The idea of the proposed architecture is to create extra feedback paths around the modulator to reduce the DAC errors further by properly designing the error cancellation logic. Transfer functions show that the DAC error at the final stage of the proposed architecture is totally cancelled, while DAC errors from other internal stages are shaped by an order higher than those in a conventional cascaded modulator. The difficulty in circuit implementation of modulators with high resolution and bandwidth increases due to the imperfection of analog components in VLSI processes. Structural and circuit-level compensation techniques are generally used in developing such modulators. Major analog nonideal effects in a multibit cascaded ÓÄ modulator include coefficient mismatches, DAC nonlinearity errors, and integrator leakages. While providing solutions for each of these nonidealities, this dissertation focuses on the minimization of the DAC error since it causes the most performance deterioration. A configurable fourth-order (2-1-1) ÓÄ modulator is implemented for architecture verification. This modulator can be configured as the proposed architecture as well as a conventional cascaded structure with various modulator orders. The design of the system's parameters and analog blocks are fully described in this dissertation. The system is fabricated by the AMI Semiconductor (AMIS) 0.5ìm double-poly triple-metal mixed- signal process through the MOSIS service. Measurement results show that with on-chip error of ±0.15 LSB for each DAC and an oversampling ratio (OSR) of 32, an improvement of 8dB of the proposed architecture over the conventional structure is observed.Item An artificial neural network for wind-induced damage potential to nonengineered buildings(Texas Tech University, 1996-12) Sandri, PraveenExtreme winds such as hurricanes and tomadoes can be extremely destmctive and result in catastrophic property losses and loss of human lives. The need to predict damage and reduce loss of life and property is becoming more important with increasing urban sprawl. Artificial Neural Networks (ANNs) provide a novel approach for representing the wind-induced damage potential prediction model. Modeled loosely after the biological neural networks of the human brain, ANNs are generally used in situations where the interactions between the input and the output variable are too complicated for an analytical solution or where there is not suflficient understanding of the problem domain. Predicting wind-induced damage potential to nonengineered buildings is not a simple task because of the complexity of constmction and limited understanding of the wind efifects on buildings. This research concentrates on the investigation of the applicability of ANNs to wind-induced damage potential prediction and the corresponding implementation issues. Even after years of post disaster windstorm damage investigations consistent, complete and robust damage information is not available to train the ANN. Thus, synthetic data instead of observed building damage information is used. WIND-RITE*, a knowledge based expert system for grading individual buildings in windstorms is used to provide the necessary damage information for the synthetic data. This research shows that a feedforward multi-layer neural network with a modified backpropagation learning algorithm can be used effectively to model wind-induced damage potential predictions for nonengineered buildings. As few as four hundred building samples are suflficient to train the network to leam the underlying relationships between the features of the building and its corresponding building damage potential. During training the ANN model is able to leam the relationships between the input features and the resulting building damage grade eflfectively. It was also found that the ANN is able to predict reasonably for samples it has not seen before.Item Applications of neural networks for distillation control(Texas Tech University, 1996-05) Munsif, Himal P.Distillation control is difificult because of its nonlinear, interactive, and nonstationary behavior; but, improved distillation control techniques can have a significant impact on improving product quality and protecting environmental resources. Advanced control strategies use a model of the process to select the desired control action. While phenomenological models have demonstrated efficient control of highly nonlinear and interactive distillation columns, they are often computationally intensive. Neural networks provide an alternate approach to modeling process behavior, and have received much attention because of their wide range of applicability, and their ability to handle complex and nonlinear problems. The main advantage in using neural network models is that they are simple, and computationally extremely efficient. In this study, neural networks were used as models in an advanced model-based control framework. Feedforward neural network models were developed using both steady-state and dynamic data to model three distiUation case studies: (i) a propylenepropane (C3) splitter; (ii) a toluene-xylene splitter; and (ui) an industrial multicomponent distillation column. Rigorous simulators were developed for these three processes which provided the data for training the networks. The neural networks were trained using a nonlinear optimization algorithm.Item Chaos and learning in recurrent neural networks(Texas Tech University, 1995-08) Corwin, Edward M.Recurrent neural networks have received a great deal of attention recently because of the variety of dynamic behaviors produced by such networks. However, attempts to train recurrent networks to produce chaotic behavior have met with great difficulty. This work examines a fundamental problem with training recurrent neural networks to produce discrete chaotic sequences and proposes a training approach which addresses the difficulties outlined. A major result of this work is a proof that a continuous model, with a bounded derivative, for a discrete chaotic system does not exist. The implication of this result is to motivate a training algorithm derived using discrete mathematics. Other algorithms assume the existence of a continuous model, which has now been shown not to exist. Networks were trained to several data sets using the traditional continuous methods and the discrete algorithm. The discrete rule improved training accuracy for all networks for the given data sets. The discrete training rule, for which one time step and multiple time step variations are presented, produced better results than Logar's training rule for the Aihara-style network and better results than Pearlmutter's algorithm for his network. A simplified proof of Hayashi's training rule, based on discrete mathematics is also presented and has the advantage of being extendible to a recursive multiple time step training algorithm. A new hybrid network, and the associated learning rules, is also presented in which coupled oscillators are positioned inside of a feed forward network. This approach alleviates some of the difiRculties inherent in training a pure oscillator network and greatly improved training accuracy. Extensions to the weight projections algorithm are also presented. The main results are that a quadratic approximation is the most eflfective choice for fitting a curve to the weight trajectory, that the extrapolation distance can be doubled if points generated later in the trajectory are more heavily weighted than those generated earlier, and that a goodness of fit test can detect a poor projection in advance of making the extrapolation.Item Cultural enhancement of neuroevolution(2002) McQuesten, Paul Herbert; Miikkulainen, RistoAny transmission of behavior from one generation to the next via non–genetic means is a process of culture. Culture provides major advantages for survival in the biological world. This dissertation develops four methods that harness the mechanisms of culture to enhance the power of neuroevolution: culling overlarge litters, mate selection by complementary competence, phenotypic diversity maintenance, and teaching offspring to respond like an elder. The methods are efficient because they operate without requiring additional fitness evaluations, and because each method addresses a different aspect of neuroevolution, they also combine smoothly. The combined system balances diversity and selection pressure, and improves performance both in terms of learning speed and solution quality in sequential decision tasks.Item Digital VLSI implementation of artificial neural network systems(Texas Tech University, 1993-05) Chamon, Jorge de. C.Not availableItem ECG time series prediction with neural networks(Texas Tech University, 1995-08) Christiansen, Brian ThomasThe comparison of three neural network methods for the prediction of a time series is studied. The digitization of electrocardiograph recordings gathered from a group of patients by the Massachusetts Institute of Technology Division of Health Sciences and Technology serve as the base for the time series to be predicted. The feed-forward back propagation learning algorithm, radial basis functions with orthogonal least squares learning algorithm and recurrent networks with Pearlmutter's learning algorithm are used as the three neural networks for prediction. The three methods prove successful in single point prediction and give fairly good results for as much as 5-point prediction, but beyond that the results are poor. The five points predicted represent less than one-quarter of a second of electrocardiograph recording time; thus showing all three methods unsuccessful as long term predictors.Item Evaluating Impulse C and multiple parallelism partitions for a low-cost reconfigurable computing system.(2009-04-01T12:08:36Z) Li Shen, Carmen C.; Duren, Russell Walker.; Engineering.; Baylor University. Dept. of Electrical and Computer Engineering.Impulse C is a C-to-HDL compiler from Impulse Accelerated Technology that facilitates the introduction of software programmers, mathematicians, and scientists, into the realm of FPGA-based algorithm development for high-speed numerical computation. This thesis evaluates the Impulse C programming language and explores differing levels of parallelism across multiple, homogeneous, FPGA development platforms using the Aurora serial communication scheme. Impulse C and Xilinx IP cores are employed in the numerical computation of a neural network consisting of 27 inputs and 1200 outputs. The artificial neural network is capable of emulating an underwater acoustic environment and has been used to determine characteristic parameters of reflections from the ocean floor. Timing, logic utilization and ease-of-use are metrics used to evaluate Impulse C in the automatic generation of VHDL code for the network test application. Implementations with parallelism at the system level and at the intermediate (loop) level are explored as part of this study.Item Evolving visibly intelligent behavior for embedded game agents(2006) Bryant, Bobby Don; Miikkulainen, RistoMachine learning has proven useful for producing solutions to various problems, including the creation of controllers for autonomous intelligent agents. However, the control requirements for an intelligent agent sometimes go beyond the simple ability to complete a task, or even to complete it efficiently: An agent must sometimes complete a task in style. For example, if an autonomous intelligent agent is embedded in a game where it is visible to human observers, and plays a role that evokes human intuitions about how that role should be fulfilled, then the agent must fulfill that role in a manner that does not dispel the illusion of intelligence for the observers. Such visibly intelligent behavior is a subset of general intelligent behavior: a subset that we must be able to provide if our methods are to be adopted by the developers of games and simulators. This dissertation continues the tradition of using neuroevolution to train artificial neural networks as controllers for agents embedded in strategy games or simulators, expanding that work to address selected issues of visibly intelligent behavior. A test environment is created and used to demonstrate that modified methods can create desirable behavioral traits such as flexibility, consistency, and adherence to a doctrine, and suppress undesirable traits such as seemingly erratic behavior and excessive predictability. These methods are designed to expand a program of work leading toward adoption of neuroevolution by the commercial gaming industry, increasing player satisfaction with their products, and perhaps helping to set AI forward as The Next Big Thing in that industry. As the capabilities of research-grade machine learning converge with the needs of the commercial gaming industry, work of this sort can be expected to expand into a broad and productive area of research into the nature of intelligence and the behavior of autonomous agents.Item Fuzzy neural networks(Texas Tech University, 1998-12) Guven, MuratSince the development of computer technology, methods have been developed and investigated to mimic the processes of the human brain. The human brain is a collection of billions of neurons interconnected with each other. Interconnected neurons are modeled with artificial neural networks (ANNs or NNs). Neural networks, mathematically speaking, are a system of linked parallel equations that are solved simultaneously and iteratively. Initial research can be found in papers by McCulloch-Pitts (1943), Hebb (1949), Rosenblatt (1958), Minsky-Papert (1969), and Hopfield (1982). Since 1982, research into neural networks has exploded and the use of neural networks to solve complex nonlinear problems has expanded (from pattem recognition to actual learning to playing games). Many different neural network architectures (the feedforward network, CMAC, Hopfield network, Kohonen network) have been developed to aid in the solution of these problems. In this paper, we are interested in the feedforward network.Item Genetic algorithms with functional mutation and mating operators in time series data mining(Texas Tech University, 2004-08) Huang, JianyongRecently, genetic algorithms (GAs) and artificial neural networks (ANNs) have been widely used in time series data mining (TSDM). Both GAs and ANNs are inspired from natural processes. A GA can be used to find optimized parameters for a given model, while an ANN has the ability to approximate unknown functions to any degree of desired accuracy without knowing the model. There are some limitations of using GAs or ANNs individually in TSDM. For example, ANNs generally use backpropagation learning algorithms, which are based on the deepest descent algorithm. Therefore, a solution from the .A.NN usually is a local optimized solution. The purpose of this thesis work is to develop innovative algorithms which can overcome the limitation of using GAs or ANNs solely in TSDM. The first part of this research involves designing a new genetic algorithm (called mGA), which can analyze not only polynomial but also non-polynomial time series. The mGA automatically searches a polynomial function with minimal degree for a non-polynomial time series. The rest of this research focuses on developing a neural network based genetic algorithm (called nGANN). The nGANN represents a chromosome as a neural network and uses genetic operators to select a global solution for a lime series. The nGANN introduces a new mating scheme (called NN _ mate), which uses a backpropagation learning network to produce offsprings. Therefore, NN mate can mate two parents with different models. The solution found by the nGANN has two attractive features: a network with small number of hidden neurons and a small mean squared error. From the solution network, h is possible to discover some relationships among different variables. Three different types of lime series data are used to evaluate the performance of the above algorithms, the two algorithms work well for one-variable polynomial and one-variable non-polynomial time series data. For two or more variables, the above algorithms do not produce very good results. In the last part of this thesis, future work is discussed.Item Inversion for explanation capability of neural networks and query-based learning(Texas Tech University, 1999-05) Saad, Emad W.Neural network inversion is the process by which we obtain the set of neural network inputs which produce a specific output. Network inversion can be used to generate an explanation of the neural network behavior. Neural networks are known for their powerful capability to model real systems by learning from examples. However, a known drawback is their "black box" character. By explanation capability we mean the expression of the knowledge learned by the neural network, in the form of comprehensible rules, so the neural network's decisions are understandable to humans. Network inversion is also the core of query-based learning (QBE). QBE is known as an active learning technique. The training data is selectively generated, such that it covers areas in the input space of high information content. This dissertation explores the use of network inversion in these two areas. Different means of inversion are presented and gradient descent inversion of the probabilistic neural network (PNN) is derived. A new technique is proposed, which generates an explanation of the neural network decision when used in classification. The proposed technique is able to generate rules with arbitrarily desired fidelity. A survey of the already existing neural network explanation algorithms is presented. Rule extraction is analyzed from an information theory point of view. The new explanation technique is applied to benchmark problems as well as to a real aerospace problem. A causality index, which provides preliminary neural network explanation, is analyzed and is applied to compare with the proposed explanation technique. QBE is applied to two real aerospace problems. The first application is a decision problem. The second application is a mapping problem with continuous output. Sigmoid scaling and jitter are explored as means of improving QBE.Item Lagrangian multiplier in the Pearlmutter algorithm and dynamic neural networks(Texas Tech University, 1998-05) Sun, JuanNot availableItem Performance characterization of higher-order neural network associative memories(Texas Tech University, 1991-05) Wang, Jung-hua.Item Phase response characterization of object similarity using the Kohonen model(Texas Tech University, 1998-05) Nath, Jagath ChandrikaNot availableItem Quantitative analyses of associative memories(Texas Tech University, 1992-05) Huang, Yo-pingNeural networks have been studied for many years in the hope of simulating human-like activities such as recognizing a friend in a picture. Associative memories are systems that can recall stored data by specifying all or a portion of a probe that has been associated or paired with that data. Until now, most of the researchers used the equal probability neuron status assumption to derive the system performance. Only a few considered the non-equally distributed case. In this dissertation, we have quantitatively analyzed the characteristics of a variety of sparsely encoded associative memories. Based on each neuron operating close to its threshold, a dynamic thresholding scheme is proposed. From this dynamic approach, the first-order sparsely encoded associative memory storage capacity is shown to have better performance than for an ordinary associative memory. Sensitivity of storage capacity with respect to variations of threshold change is calculated to observe the effect on capacity change. Information capacity is also investigated in order to choose the optimum activity rate. Extensions are made to the higher order system. Several properties such as storage and information capacities are explored to evaluate the system performance. Other contributions include: (1) consideration of the retrieval problem of stored patterns in the noisy environment, and (2) development of the fault tolerant analysis of associative memories. Both neuron and connection fault models are analyzed in detail. Simulation results are shown to be consistent with theoretical work.Item Recognition of Alzheimer's disease using quantitative electroencephalography(Texas Tech University, 2004-08) Manion, Robert VincentCurrently, Alzheimer's disease is diagnosed through a lengthy process, including patient history, neuropsychological testing, neurophysiological analysis, and psychological evaluations. There is hope that a quantitative, objective diagnostic procedure would increase diagnosis capabilities, including earlier detection, an increase in ease of diagnosis, and greater diagnosis consistency. This thesis investigates using power values and complexity measures of the electroencephalogram as input features to a neural network for classification of Alzheimer's disease patients, mild cognitively impaired patients, and control subjects. Specifically, the complexity measures activity, complexity, and mobility, as well as the relative power values in frequency bands delta, theta, alpha, beta, and gamma are calculated for the electroencephalogram of each subject. Finally, a Learning Vector Quantization Neural Network will be designed to classify each subject into their respective category based upon an input vector consisting of these power features. The ability of this network to classify patients correctly will be measured and reported.Item Recognition of handwritten letters using a locally connected back-propagation neural network(Texas Tech University, 1991-05) Gomez-Gil, Maria del Pilar.Item Recognition of patterns in electronic communication signals using neural networks(Texas Tech University, 1991-05) Stubbendieck, Gregg T.This paper presents the results of research into automatic recognition of a class of electronic communication signals using a Back Propagation (BP) model neural network. Communication signals present an important and interesting pattern recognition challenge since they change unpredictably over time in accordance with the information they carry. There are situations in which a receiver has no prior knowledge of a particular signal and must classify it before interpreting it. The communication systems of interest here use frequency division techniques to multiplex several telegraph sub-signals in a standard communication channel. Previous research in recognizing these signals has demonstrated good recognition rates at the cost of expensive signal preprocessing. In this research, a BP network, smaller than networks used previously on this problem, was trained to recognize several types of these signals with a high degree of accuracy using a feature vector that is computationally less expensive and smaller than previous feature vectors. The observation that the BP network is tolerant of noise in patterns is reaffirmed in this research.