Browsing by Author "Steve G Romanuik"

Now showing 1 - 6 of 6
  • No Thumbnail Available
    Item
    Fuzzy Rule Extraction for Determining Creditworthiness of Credit Applicants
    (1992-10-01T00:00:00Z) Steve G Romanuik
    The main objective of this research paper is to provide an empirical analysis of the hybrid symbolic/connectionist expert system development tool SC-net to act as a viable system for acquiring expert system knowledge by means of learning. The task to be studied is the prediction of creditworthiness for credit seeking applicants. The creditworthiness domain - unlike many other domains studied by the machine learning community - contains both uncertainties in the inputs and outputs. Apart from showing SC-net's ability to derive human acceptable models for this data, strong emphasis is placed on deriving rules that can adequately describe the imprecision inherent in such domains. No a priori domain knowledge, such as pre-defined fuzzy membership functions or pre-selection of important input features is required. The affect of training set size on number of rules and attributes per rule is addressed and a sample set of extracted rules with derived membership functions is provided. In all cases acceptable models for determining creditworthiness are derived. The herein described experimental results should further strengthen SC-net's ability to act as a knowledge acquisition tool for obtaining acceptable expert knowledge in uncertain domains.
  • No Thumbnail Available
    Item
    Multi-Pass Instance Based Learning
    (1993-03-01T00:00:00Z) Steve G Romanuik
    This paper introduces a new modified approach to the instance based learning theory. Instance based learning is augmented by neighborhood spheres and multi-pass training to improve both on generalization capabilities and storage requirements. Two models for creating neighborhood spheres are investigated and put in perspective with the IBL instance based learner. The IBL system considered here is based on the proximity algorithm, the growth (additive) algorithm and a noise resistant modification of the growth additive) algorithm. The herein described experiments will address the similarity of the MPIL and the IBL algorithms, but also point out significant differences in the approach of reducing storage requirements and increasing generalization. A time complexity analysis of the proposed multi-pass instance based learning approach is provided. Several domains are used in this study, which include a real world domain in CMOS wafer fault diagnosis to allow for a comparison of these two approaches. Finally, the task of knowledge extraction in form of rules is addressed.
  • No Thumbnail Available
    Item
    Prove of Convergence of Extended Divide & Conquer Networks
    (1993-04-01T00:00:00Z) Steve G Romanuik
    The task of determining an effective architecture for multi-layer feed forward backpropagation like neural networks can be a time consuming effort. Over the past couple years several algorithms were proposed for dynamically constructing network architectures. Some of these algorithms have been shown to converge for binary data. Unfortunately, the results do not carry over for the non-binary input case. For classificatory problems no guarantees are provided, which would suggest otherwise. In this paper, we present an extension to the basic network growing algorithms, which allows constructing networks in bounded time. The algorithm guarantees convergence for any classificatory domain, as long as no contradictory training examples are present. The extension is described for the Divide \& Conquer Networks (DCN) algorithm. The derived mathematical model can be readily incorporated into other network growing approaches to ensure convergence. The model is dependent on the usage of simple threshold cells, which can be applied to a variety of learning rules.
  • No Thumbnail Available
    Item
    Representing Complex Fuzzy Membership Functions in a Connectionist Network
    (1992-12-01T00:00:00Z) Steve G Romanuik
    The problem of deriving membership functions as a means for describing linguistic variables (for some control process) and the choice of fuzzy inference operators and connectives is at the heart of developing fuzzy control systems. Over the years connectionist systems have obtained prominence as a means to solve complicated learning tasks. More recently a surge in interest for applying neural systems to fuzzy control problems has occurred. In this paper, it will be shown how complicated fuzzy membership functions can be composed of simpler pi-shaped functions. The importance lies in the fact, that the decomposition process can be implemented through use of a connectionist representation. Furthermore, it allows augmenting an existing hybrid symbolic/connectionist learning system - SC-net -to incorporate fuzzy variables described by more complicated membership functions than was earlier possible. Finally, it provides the necessary tools for constructing connectionist fuzzy controllers, which by means of machine learning can be trained to adapt to an ever changing environment.
  • No Thumbnail Available
    Item
    Theoretical Analysis of Hybrid Symbolic/ Connectionist Networks
    (1992-10-01T00:00:00Z) Steve G Romanuik
    Abstract not available.
  • No Thumbnail Available
    Item
    Theoretical Results for Applying Neural Networks to Lossless Image Compression
    (1994-03-01T00:00:00Z) Steve G Romanuik
    The ability to employ neural networks to the task of image compression has been pointed out in recent research. The pre-dominant approach to image compression is centered around the backpropagation algorithm to train on overlapping frames of the original picture. Several deficiencies can be identified with this approach: First, no potential time bounds are provided for compressing images. Second, utilizing backpropagation is difficult due to its computational complexity. To overcome these shortcomings we propose a different approach by concentrating on a general class of 3-layer neural networks of 2(N+1) hidden units. It will be shown that the class ${\cal N}^{*}$ can uniquely represent a large number of images, in fact, growth of this class is larger than exponential. Instead of training a network, it is automatically constructed. The construction process can be accomplished in ${\cal O}_{Worst}(n) = n^{4} - n^{2}$ time, where $n$ is the image size. Obtainable compression rates (lossless) exceed 97\% for square images of size 256.