HomeChair: Ling Guan, University of Sydney, Australia
HomeJong-Min Park, University of Wisconsin-Madison (U.S.A.)
Reported in this paper is an intelligent agent that aids users to conduct efficient Internet Web information retrieval through query formulation, information collection, information clustering, and analysis. The underlying mechanism is a probabilistic Sample-at-the-boundary learning algorithm for clustering the search results and learning and matching the user concept. Kohonen's ``windowed'' Learning Vector Quantization algorithm is shown to be related to this Sample-at-the-boundary learning algorithm. A prototype system has been developed and evaluation has been conducted.
ic981429.pdf (From Postscript)
TOP
Kaare J Jensen, Technical University of Denmark (Denmark)
Steen M Munk, NESA A/S (Denmark)
John A Sorensen, Technical University of Denmark (Denmark)
A new approach to the localization of high impedance ground faults in compensated radial power distributionnetworks is presented. The total size of such networks is often very large and a major part of the monitoringof these is carried out manually. The increasing complexity of industrial processes and communication systemslead to demands for improved monitoring of power distribution networks so that the quality of power deliverycan be kept at a controlled level.The ground fault localization method for each feeder in a network, is based on the centralized frequencybroad band measurement of three phase voltages and currents. The method consists of a feature extractor,based on a grid description of the feeder by impulse responses, and a neural network for ground faultlocalization. The emphasis of this paper is the feature extractor, and the detection of the time instanceof a ground fault.
ic981746.pdf (From Postscript) TOP
Michael Becker, University of Bonn (Germany)
Mikio Braun, University of Bonn (Germany)
Rolf Eckmiller, University of Bonn (Germany)
A tuning method with reinforcement learning (RL) for the Retina Encoder (RE) of a Retina Implant (RI) as a visual prosthesis for blind subjects with retinal degenerations is proposed. RE simulates retinal information processing in real time by means of spatio-temporal receptive field (RF) filters and generates electrical signals for stimulation of several hundreds of ganglion cells (GC) to regain a modest amount of vision. For each contacted GC, RE has to be optimized with regard to the patient's perception. The patient's (for the present simulated) evaluative feedback is applied here in a dialog module as a reinforcement signal to train several RL agents in a neural network learning process (see also http://www.nero.uni-bonn.de).
ic981964.pdf (From Postscript) TOP
Graham C Freeland, University of Strathclyde, Scotland (U.K.)
Tariq S Durrani, University of Strathclyde, Scotland (U.K.)
Deterministic multiscale defined representational forms have found a significant role in the theory and application of signalprocessing over the last decade. With little argument the most widely important for signal and system modelling is likely to be multiscale defined wavelets. Another class of multiscale representation which has attracted consistent interest over the same period is the group of signal models defined in terms of Iterated Function Systems (IFS). This paper is concerned with widening the IFS application to include system modelling, particularly of neural network-like structures. We introduce an interpolating IFS model as a form of self-organising map with global fractal constraints. Symbolic addressing is employed to discretize the attractor into pseudo network nodes. We present in detail an online gradient based algorithm for training. This particular model is intended for efficient pattern recognition in complex environments, for example, with multifractal sources such as those seen in network traffic and general turbulence.
ic982100.pdf (From Postscript) TOP
Wing-kai Lam, The Chinese University of Hong Kong (Hong Kong)
Lei Xu, The Chinese University of Hong Kong (Hong Kong)
Optimizing the number of hidden units in feedforward neural networks is an important issue in learning. Recently, a new criteria on selecting the number of hidden units in feedforward neual networks is proposed by one of the present author, based on the so-called Bayesian Ying-Yang (BYY) learning theory. The new criteria can be simply computed during the implementation of backpropagation training. In this paper, the criteria is experimentally studied and compared with the well-known Cross Validation approach. Simulation results show that obtained number of hidden units by the BYY criteria is highly consistent to the minimal generalization error and outperforms the Cross Validation approach.
ic982372.pdf (From Postscript) TOP