|
ZAPTRON Systems, Inc.
Neurofuzzy: A Different Type of Neural Nets
-- Zaptron's High-Order Nonlinear Neural Networks
|Fuzzy logic|Neural
Networks|Neurofuzzy|
|High-order Nonlinear Neural Networks|DataX|
 |
Fuzzy Logic
What is fuzzy logic - Fuzzy inference (or simply "fuzzy logic")
is a powerful problem-solving methodology with wide applications in industrial control and
information processing. It provides a simple way to draw definite conclusions from vague,
ambiguous or imprecise information. It resembles human decision making with its ability to
work from approximate data and find precise solutions. Unlike classical logic which requires a deep understanding of a
system, exact equations and precise numeric values, fuzzy logic incorporates an
alternative way of thinking, which allows modeling complex systems using a higher level of
abstraction originating from our knowledge and experience. Fuzzy Logic allows expressing
this knowledge with subjective concepts such as "very good" and "a little
bit satisfied" which are mapped into exact numeric ranges.
How does fuzzy logic work - it uses 3
simple steps defined below
 |
fuzzification - to convert numeric data (e.g., $24.50 in sales) in
real-world domain to fuzzy numbers in fuzzy domain
|
|
aggregation (rule firing) - computation of fuzzy numbers (all
between 0.0 and 1.0) in fuzzy domain
|
|
defuzzification - convert the obtained fuzzy number back to the
numeric data (e.g. 150.34% in total profitability) in the real-world domain.
|
Fuzzy
logic advantages:
|
mimic human decision making to handle vague concepts
|
|
rapid computation due to intrinsic parallel processing nature
|
|
ability to deal with imprecise or imperfect information
|
|
resolving conflicts by collaboration, propagation and aggregation
|
|
improved knowledge representation and uncertainty reasoning
|
|
modeling of complex, non-linear problems |
|
natural language processing/programming capability
|
Fuzzy logic limitations:
|
highly abstract and heuristic |
|
need experts for rule discovery (data relationships)
|
|
lack of self-organizing & self-tuning mechanisms of NN |
|
|
Neural Networks (NN)
Neural
networks are modeless systems that learn from the underlying relationships of data.
They are organized in a way to simulate the cells of human brain.
|
|
Neurofuzzy Techniques
Neurofuzzy refers to the combination of fuzzy set theory
and neural networks with the advantages of both:
|
handle any kind of information (numeric, linguistic, logical, etc.)
|
|
manage imprecise, partial, vague or imperfect information |
|
resolve conflicts by collaboration and aggregation |
|
self-learning, self-organizing and self-tuning capabilities
|
|
no need of prior knowledge of relationships of data |
|
mimic human decision making process |
|
fast computation using fuzzy number operations |
|
|
ZAPTRON's Proprietary Technology
-- High-Order Nonlinear Neural Networks.
Unlike classical, linear neural networks, the proprietary High-Order Nonlinear
Neural Networks (HONLNN) by ZAPTRON have the following outstanding features:
|
Self-organizing ability: HONLNN traces model changes or variations
(such as business rule, relationship or regulation changes). For given data the HONLNN
system can automatically organize itself as an optimal model for the data, without human
interaction. This means, the number of hidden layers and the number of nodes on each
layer are determined by the HONLNN itself, a significant improvement over the backward
propagation (BP) networks.
|
|
Nonlinearity - the inter-node effect is added to each
node's output, giving better approximation than that by linear networks.
|
|
Modeling of fuzzy information - fuzzy set theoretical method is
incorporated in HONLNN to allow modeling of quantitative information expressed by
linguistic or logical data, such as human, natural or environmental factors in finance and
business management.
|
|
Fusion of incompatible information - HONLNN uses fuzzy log
to offer a way for combining numeric data with linguistic or logical data to achieve
information propagation and knowledge aggregation.
|
|
Super-fast computation: input data are first transferred by a
high-order nonlinear module before being used in NN learning. By the equivalence of
NN and Fourier series, a super-fast, real-time NN is thus produced.
|
|
Incorporating GMDH (Group Method for Data Handling): GMDH is a
high-order, nonlinear method in system model identification. Like Fourier series,
GMDH can approximate any function with good accuracy.
|
|
Genetic algorithm-based
generation - network layers and nodes are generated iteratively by genetic programming,
giving the best solution.
|
|
Achieve global minimum quickly by using fewer layers
|
|
|
Applications
of Neurofuzzy Technology
|
Business rule extraction and explanation process |
|
Incorporation of personal preference |
|
Model building |
|
Fusion of numeric data and linguistic information
|
|
|
Tools in DataX™ Software Suite
|
Build business models |
|
Market simulations |
|
Model Validations |
|
Fuzzy time series forecasting |
|
|
