Massey Documents by Type

Permanent URI for this communityhttps://mro.massey.ac.nz/handle/10179/294

Browse

Subject: search.filters.subject.Neural networks (Computer Science)

Search Results

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    A neural network based window filter and its training for image processing tasks : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Technology at Massey University, Palmerston North, New Zealand
    (Massey University, 1995) Pugmire, Ralph Harold
    The design and implementation of a neural network based universal window filter (NNWF) is described. Experiments are reported which demonstrate that such a network filter can learn to perform filtering operations from input-target image pairs. Difficulties with training such a filter for more complex tasks are then described. Standard methods of improving the learning performance of neural networks are reviewed. Speculation on development of intelligent systems is presented with particular reference to the purpose and use of logical sequential thought and rules. The importance of an educational environment is outlined and a series of four heuristics for improving the training of neural networks is suggested. Initial experiments with these heuristics are described. The analogy to instructional design theory for humans is proposed and a formal basis for the design of an educational environment for the neural network based window filter is developed. Finally, a series of experiments are reported which test the validity of the use of the educational environment and demonstrate the effectiveness of the methods developed for implementing such an environment for the NNWF.
  • Thumbnail Image
    Item
    Equity trend prediction with neural networks : an empirical analysis : a thesis presented in partial fulfilment of the requirements for the degree of Master of Engineering in Computer Systems Engineering at Massey University, Albany, New Zealand
    (Massey University, 2012) Halliday, Russell James
    This thesis presents results of neural network based trend prediction for equity markets. Despite a breadth of research which has focused on the prediction of various equity and currency exchange markets, much has focused on the use of specific techniques in such predictions. Few bodies of work have compared a wide range of equity market data preprocessing and technical analysis techniques in creating a prediction model based on feed-forward and recursive neural networks. To achieve a broad-based prediction model, the work in this study was broken into three distinct parts. Firstly, the neural networks goal is defined as finding whether a stock will be higher or lower than the previous trading period. Subsequent to this, a variety of input data scaling and network topologies are looked at. This includes the use of Self Organising Maps (SOM) as a data classification method to limit neural network inputs and training data requirements. Feed Forward and Elman networks of various topologies are used to narrow down the best network combinations. The resulting simulation is a neural network that can predict whether the next trading period will be, on average, higher or lower than the current. Secondly, the topology and preprocessing lessons learned during the first phase are applied to two types of neural network. Technical analysis is applied to the input data in an attempt to verify the usefulness of conventional stock indicators as inputs to neural networks. The two types of networks trained are, for the purposes of this thesis, dubbed indicator-predictive and price-predictive networks, meaning that technical analysis inputs are used to predict the next trading days technical indicator or future stock price direction respectively. Finally, a combined network is trained which takes the inputs from the price-predictive networks in an attempt to gain better results. The hypothesis with this network is that the combined neural network should learn which of its inputs are more indicative of a stock price movement, and thus more accurately predict the future direction of the stock.
  • Thumbnail Image
    Item
    The effects of using problem knowledge in a neural network for image processing tasks : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Computer Science at Massey University, Palmerston North, New Zealand
    (Massey University, 2001) Gunetileke, Kapila Sanjeeva
    This thesis is concerned with aspects of computational intelligence. Computational intelligence is a new paradigm of artificial intelligence based on biological intelligence. Computational Intelligence explores the potential for biologically inspired intelligent adaptive machines and behavior. A relatively new and important sub discipline within the field of computational intelligence, is that of neural networks. Neural networks are networks of artificial neurons with a high degree of interconnectivity. The networks capture and accumulate knowledge as the pattern of weights in the interconnections between neurons. Neural networks are iteratively trained to perform tasks by "learning" from examples. These networks are often slow to train. One of the reasons for this is that the starting weights of the network are conventionally not related to the problem being solved. In this thesis a methodology is explored that maps problem knowledge to the starting weights of a fuzzy neural network window filter (FuNNWF). The FuNNWF architecture was developed from the combination of fuzzy logic and artificial neural networks for image processing tasks. The effects of the use of problem knowledge on FuNNWF training are investigated. The problem knowledge mapping procedure is extended from boolean rule mapping to conditional rule mapping, which allows better representation of the problems. Four real world image processing problems are investigated using the new weight initialization methodology. The experiments reported in this thesis demonstrate that the use of problem knowledge improves the robustness and convergence of the neural network. It is also shown that the use of the methodology is most effective on network training when the training data is noisy, unreliable and ambiguous.