next up previous
Next: Introduction



Supervisory committee: Chairperson: professor Toshio OKAMOTO

Member: professor Mamoru HOSHI

Member: assistant professor Shunichi TANO

Member: professor Takeshi KAMBARA

Member: professor Kunikatsu TAKASE

Member: professor Etsuji TOMITA

Abstract :

The goal of this research is to build an integrated environment to serve as a possible assistant for a teacher during the educational process. Getting useful information for the teaching process is very difficult when dealing with unstructured knowledge. The complete Subsymbolic Knowledge Extraction Environment (SSKEE) system contains three building blocks: the neural network module, the knowledge extraction module and the user interface module. As the system was built from scratch, during the design and implementation many contributions and results were obtained, as follows. The first module described is the neural network module. As a building block for the neural network I used an original energy function. In this thesis I will also present the construction and usage of the energy function for supervised learning on feedforward networks, based on restrictions. The focus is on the mathematical deductions of the energy function, based on the Lyapunov (also called infinite) norm, from error minimization procedures. I will show how the movement equations derived from this energy function improve the learning and generalization capacity of the neural tool in the case of stock exchange (SE) prediction, in the sense of time-series (TS) prediction. I will also show some comparative results of my method and the classical backpropagation (BP) method, obtained by means of the T (Theill) test and the correlation computation. The verification of the proposed energy function is done through computer simulation. Moreover, in the context of the ANN implementation, the parallel aspects of an ANN, as well as the advantages of a parallel implementation of a ANN will be discussed. The different levels of parallelism, namely, coarse, medium and fine grain parallelism are put in correspondence with the ANN functions and structure. For instance, at a coarse grain parallelism, the layer parallelism is considered. At the medium level, I studied the neuron level parallelism. At the fine level I had to consider the very functions of the ANN and to enter the instruction level parallelism. Out of these types, an original combined optimization is built and described. The second module referred to in this thesis is the knowledge extraction module. Neural networks can store sub-symbolic knowledge, but until recently it was believed to be only in a "black-box" format. Knowledge extraction from ANNs is a relatively new field, which tries to reduce these disadvantages and build a bridge between sub-symbolic and symbolic knowledge. As the teaching process requires only symbolic knowledge, I believe this new tool to be a chance for teachers to significantly improve their teaching materials and/or style by combining the symbolic knowledge of the domain theory with the rules extracted from the empirical sub-symbolic knowledge stored in ANNs trained on examples. The third module, the user interface module is the one to make the link to the end user, in this case, the professor/teacher. The whole system adds up to a Neural Network's Sub-Symbolic Knowledge Eliciting Environment for a possible application in the Education Process, based on a Time-Series prediction parallel Neural Network, verified on a case study of teaching stock exchange developments in a classroom teaching process, with the theme Economy.

next up previous
Next: Introduction

Alexandra Cristea
Tue Feb 9 20:20:27 JST 1999