No description available.

No description available.

The overall objective of this research was to investigate the feasibility of using artificial neural networks to detect the incidence of clinical bovine mastitis and to determine the major factors influencing it. The first part of this research was devoted to a general examination of the learning ability of artificial neural networks by training them with relatively small data sets. These data sets (a total of 460,474 records) contained suspected indicators of mastitis such as milk production, stage of lactation and somatic cell count, and it was hoped that artificial neural networks would be able to detect what statistical modelling had already established elsewhere in the literature. The second part of this research was extended to examine the roles of more information resources such as conformation traits and their genetic values---factors that have not been studied extensively, with either conventional approaches or emerging technologies like artificial neural networks. (Abstract shortened by UMI.)

No description available.

This thesis discusses aspects of design, implementation and theory of expert systems, which have been constructed in a novel way using techniques derived from several existing areas of Artificial Intelligence research. In particular, it examines the philosophical and technical aspects of combining techniques derived from the traditional rule-based methods for knowledge representation, with others taken from connectionist (more commonly described as Artificial Neural Network) approaches, into one homogenous architecture. Several issues of viability have been addressed, in particular why an increase in system complexity should be warranted. The kind of gain that can be achieved by such hybrid systems in terms of their applicability to general problem solving and ability to continue working in the presence of noise, are discussed. The first aim of this work has been to assess the potential benefits of building systems from modular components, each of which is constructed using different internal architectures. The objective has been to progress the state of knowledge of the operational capabilities of a specific system. A hybrid architecture containing multiple neural nets and a rule-based system has been designed, implemented and analysed. In the course of, and as an aid to the development of the system, an extensive simulation work-bench has been constructed. The overall system, despite its increased internal complexity provides many benefits including ease of construction and improved noise tolerance, combined with explanation facilities. In terms of undesirable features inherited from the parent techniques the losses are low. The project has proved successful in its stated aims and has succeeded in contributing a working hybrid system model and experimental results derived from the comparison of this new approach with the two, primary, existing techniques.

Image compression using neural networks in the past has focused on just reducing the number of bytes that had to be stored even thought the bytes had no meaning. This study looks at a new process that reduces the number of bytes stored but also maintains meaning behind the bytes. The bytes of the compressed image will correspond to parameters of an existing graphic algorithm. After a brief review of common neural networks and graphic algorithms, the back propagation neural network was chosen to be tested for this new process. Both three layer and four layer networks were tested. The four layer network was used in further tests because of its improved response compared to the three layer network. Two different training sets were used, a normal training set which was small and an extended version which included extreme value sets. These two training sets were shown to the neural network in two forms. The first was the raw format with no preprocessing. The second form used a Fast Fourier Transform to preprocess the data in an effort to distribute the image data throughout the image plane. The neural network’s response was good on images that it was trained on but responded poorly to new images that were not used in the training sets. / Department of Computer Science

Artificial neural networks (ANNs) have been used in many fault detection and identification (FDI) applications due to their pattern recognition abilities. In this study, two ANNs, a supervised network based on Backpropagation (BP) learning and an unsupervised network based on Adaptive Resonance Theory (ART-2A), were tested for FDI on an automated assembly machine and compared to a conventional rule-based method. Three greyscale sensors and two redundant limit switches were used as cost-effective sensors to monitor the machine's operating condition. To test each method, sensor data were collected while the machine operated under normal conditions, as well as 10 fault conditions. Features were selected from the raw sensor data to create data sets for training and testing. The performance of the methods was evaluated with respect to their ability to detect and identify known, unknown and multiple faults. Their modelling and computational requirements were also considered as performance measures. Results showed that all three methods were able to achieve perfect classification with the test data sets; however, the BP method could not classify unknown or multiple faults. In all cases, the performance depended on careful tuning of each method’s parameters. The BP method required an ideal number of neurons in the hidden layer and good initialization. The ART-2A method required tuning of its classification parameter. The rule-based method required tuning of its thresholds. Although it was found that the rule-based system required more effort to set up, it was judged to be more useful when unknown or multiple faults were present. The ART-2A network created new outputs for these conditions, but it could not give any more information as to what the new fault was. By contrast, the rule-based method was able to generate symptoms that clearly identified the unknown and multiple fault conditions. Thus, the rule-based method was judged to be the best overall method for this type of application. It is recommended that future work examine the application of computer vision-based techniques to FDI with the assembly machine. The results from this study, using cost-effective sensors, could then be used as a performance benchmark for image-based sensors. / Thesis (Master, Mechanical and Materials Engineering) -- Queen's University, 2014-05-16 17:21:13.676

No description available.

We investigate five different problems in the field of the statistical mechanics of neural networks. The first three problems involve attractor neural networks that optimise particular cost functions for storage of static memories as attractors of the neural dynamics. We study the effects of replica symmetry breaking (RSB) and attempt to find algorithms that will produce the optimal network if error-free storage is impossible. For the Gardner-Derrida network we show that full RSB is necessary for an exact solution everywhere above saturation. We also show that, no matter what the cost function that is optimised, if the distribution of stabilities has a gap then the Parisi replica ansatz that has been made is unstable. For the noise-optimal network we find a continuous transition to replica symmetry breaking at the AT line, in line with previous studies of RSB for different networks. The change to RSBl improves the agreement between "experimental" and theoretical calculations of the local stability distribution ρ(λ) significantly. The effect on observables is smaller. We show that if the network is presented with a training set which has been generated from a set of prototypes by some noisy rule, but neither the noise level nor the prototypes are known, then the perceptron algorithm is the best initial choice to produce a network that will generalise well. If additional information is available more sophisticated algorithms will be faster and give a smaller generalisation error. The remaining problems deal with attractor neural networks with separable interaction matrices which can be used (under parallel dynamics) to store sequences of patterns without the need for time delays. We look at the effects of correlations on a singlesequence network, and numerically investigate the storage capacity of a network storing an extensive number of patterns in such sequences. When correlations are implemented along with a term in the interaction matrix designed to suppress some of the effects of those correlations, the competition between the two produces a rich range of behaviour. Contrary to expectations, increasing the correlations and the operating temperature proves capable of improving the sequenceprocessing behaviour of the network. Finally, we demonstrate that a network storing a large number of sequences of patterns using a Hebb-like rule can store approximately twice as many patterns as the network trained with the Hebb rule to store individual patterns.

In recent years financial markets have become complex environments that continuously change and they change quickly. The strong link between the continuous change in the markets and the danger of losing money when trading in them, has made financial studies a domain that concentrates increasing scientific and business attention. In this context, the development of computational techniques that can monitor recent financial events can process them according to their similarity with historical data recordings, and can support financial decision making, is a challenging problem. In this work, the principal idea for tackling this problem is the integration of 'current' market information as derived from the market's recent past and historical information. A robust technique which is based on flexible pattern matching, segmented data representations, time warping, and time series embedding dimension measures is proposed. Complementary time series derived features, concerning trend structures, temporal considerations and statistical measures are systematically combined in this technique. All these components have been integrated into a software package, which I called PROGNOSIS, that can selectively monitor its application and allows systematic evaluation in terms of financial forecasting and trading performance. In addition, two other topics are discussed in this thesis. Firstly, in chapter 3, a neural network, that is known as the Growing Neural Gas network, is employed for financial forecasting and trading. To my knowledge, this network has never been applied before to financial problems. Based on this a neural network forecasting and trading benchmark was constructed for comparison purposes. Secondly, a novel method of approaching the well established co-integraton theory is proposed in the last chapter of the thesis. This method enhances the co-integration theory by integrating into it local time relations between two time series. These local time dependencies are identified using dynamic time warping. The hypothesis that is tested is that local time shifts, delays, shrinks or stretches, if identified, may help to reveal co-integrating movement between the two time series. I called this type of co-integration time-warped co-integration. To this end, the time-warped co-integration framework is presented as an error correction model and it is tested on arbitrage trading opportunities within PROGNOSIS.