Neural Networks and Fuzzy Logic Projects
Scope
Links within Australia are with the University of New South Wales: Schools of
Petroleum Engineering, Computer Science and Engineering, Civil Engineering.
The University of Queensland Department of Computer
Science; the University of WA: Department of Computer Science, and the
Centre for Intelligent Information Processing Systems in the Department of
Electrical and Electronic Engineering.
International links are with researchers in Hungary, Japan, Singapore, and
the USA.
The connecting themes between the projects described below are: the use of
neural network and fuzzy logic techniques in practical applications from a
strong theoretical basis; and the development of techniques to simplify the
building of successful neural network and fuzzy logic applications. Financial
support for these activities has been received from AAECP, AIDAB, ATERB, ARC
(including 2 current large ARC grants) and Nortel Networks.
Short Descriptions of Project Areas
Extracting Rules and Explanations from trained neural networks
TD Gedeon, H Turner, ECY Choi
From any program or person who is not a mythical oracle, any advice or
conclusion must be explainable. Neural networks can learn to reproduce expert
conclusions and categorisations in difficult domains where no clear rules are
known. Our methodology uses the causal correlation of inputs and outputs to
produce explanations which include: a set of rules which a particular case
satisfies or distinguishes from other similar cases; the significant (and
sufficient) features of the case; the features which if changed would lead to
different conclusions; the most similar archetypal case; and the next most
likely decision.
This method has been used in applications such as developing symbolic models of
the Gross Domestic Product in third world nations, and the prediction of final
marks in undergraduate Computer Science subjects.
Hybrid systems - Neural networks and Expert Systems
TD Gedeon, LK Milne, RP Good
Connectionist techniques (neural networks) are better for some tasks such as
perception involving the fuzziness of the real world, while symbolic techniques
(expert systems) are superior for other tasks involving the world of the mind.
Many real problems are a mixture of these two extremes requiring a combination
of the strengths of these two approaches.
Some work has been done in Geographic Information Systems in the development of
hybrid models for predicting forest resources using remotely sensed and
ancillary spatial data. Other work is in the interactive modelling of Gross
Domestic Product in developing countries, which also involves hybrids of
different neural network models.
Neural Networks for Genetic Petroleum Reservoir Classification and
Well Log Interpretation
PM Wong, TD Gedeon, IJ Taggart
Accurate estimation of porosity and permeability from wireline logs is a
central problem in log interpretation. The quality of these values has a
strong impact on reservoir modelling and the subsequent management scheme.
Recent studies have shown that identification of lithofacies prior to
estimating hydraulic properties can lead to improved characterisation.
This project examines the use of neural and neuro-fuzzy networks in the
genetic classification of rock types in a complex depositional environment,
and the estimation of porosity and permeability using well log data.
This work is being extended to use fuzzy and knowledge based techniques to
capture expert knowledge automatically. This is to be used in the automated
inferencing of petro-physical properties of oil-bearing rock formations for
reservoir classification using seismic and well log data.
Classification of schizophrenics using Eye gaze data
TD Gedeon, E Gordon, K Li, K Kozek
Differences in visual behaviour of schizophrenics when looking at
particularly face images showing some emotion can be detected. This
work has used neural network and hidden markov models to distinguish
schizophrenics from controls using the output of an infra-red eye
gaze monitor.
Interpolation in sparse fuzzy rule bases
TD Gedeon, LT Koczy, P Baranyi
There are a number of methods for generating fuzzy conclusions for
observations in comprehensive fuzzy rule bases. Sometimes there are
natural gaps, however, in the knowledge base. Also, sparse rule bases
which do not contain redundant information can offer some computational
advantages.
Fuzzy rule interpolation is used to provide conclusions for observations
for which there may be no overlap even with the supports of the existing
rules in the rule base.
In this project we are developing fuzzy interpolation techniques which are
conservative with respect to the degree of local fuzzyness of the rule-base,
form intuitively acceptable conclusions, and have low computational complexity.
Automated inferencing for petroleum reservoir properties
TD Gedeon, PM Wong, A Ramer
The flow capability of sedimentary rocks (permeability) in petroleum
reservoirs directly affects hydrocarbon production and the definition of
reserves. The determination of permeability throughout a reservoir is a
complex problem because laboratory results on rock samples are only
available in limited and isolated well locations. The laboratory reports
contain hard data (measured permeability values) as well as soft data
(qualitative or linguistic statements). We are developing techniques to
integrate these two kinds of data to produce expert aides for reservoir
modelling.
Analytical models and the complexity of fuzzy rule bases
LT Koczy, A Zorat, TD Gedeon
Approximate models using fuzzy rule-bases can be made more precise by
suitably increasing the size of the rule-base. A larger rule-base, however,
requires more time for its evaluation and hence the problem arises of
determining a size that is "good enough" for the task at hand.
This trade-off
between computation time and precision is considered in this project in the
context of tracking a moving target. In this problem a higher degree of
accuracy results in a tighter precision of the target location, but at the
cost of longer computation time, during which the target can move further
away, thus ultimately requiring a longer search time for exact target
location.
Ambiguous information architecture
TD Gedeon, LA Coward, B Zhang, N Treadgold
A design approach in which functional components can exchange ambiguous
information leads to systems with the recommendation architecture which
are less subject to the limitations of traditional functional architectures.
These are the difficulty in the design of systems to perform parallel
processing, and extreme difficulty in design of systems which can
heuristically change their own functionality.
Biological brains have been constrained by natural pressures to adopt
functional architectures with this different information exchange approach.
This work uses a biologically inspired neural network architecture,
for efficient IP network and transport network partitioning.
Information Compression in Fuzzy Systems by Interpolative Approaches
TD Gedeon, LT Koczy
Fuzzy expert systems built with linguistic rules (which reflect
uncertainty, or vagueness of concepts in natural language) have very wide
potential application where human expertise is difficult to express in
crisp mathematical form. For real problems many rules are required, which
takes too much computing time to solve any but the simplest problems. We
are developing techniques to reduce the number of rules required, and to
partition them for more efficient computation. This will lead to much
wider applications of fuzzy systems, and reduce computation costs
significantly.
Contact
T.D. Gedeon
- Ph +61 2 6125 1052, Fax +61 2 6125 0010, tom@cs.anu.edu.au
Wednesday June 30 18:42:16 WST 1999