There are several software packages, which implement GMDH algorithms as a further
development of software from Glushkov Institute of Cybernetics
and ASPN-II (Algorithm for Synthesis of Polynomial Networks)
from Barron Associates Co., USA.
GMDH brings high-end modeling capabilities on desktop without
the need of being an expert in modeling since it will learn completely
unknown relationships between outputs and inputs of any given system
in an evolutionary way from a very simple organization to an optimal
complex one by itself. The main advantages of software which realize
inductive approach are:
- Only minimal, uncertain a priori information about the system
is required. That means, even if you are not an expert in modeling,
data analysis or designing a neural network you will be able to
model, analyse and predict very complex objects of nearly any
kind of system.
- Very fast and effective learning process for ordinary PC's.
That means, you can solve problems on your desktop in a reasonable
time which you may have never thought about before.
- Modeling on very short and noisy data samples. That means, you
can deal with a problem as is and don't have to construct artificial
conditions for your modeling method to get it work.
- Output of an optimal complex and cross-validated model. That
means, you commonly can expect to get a model which is robust,
as simple as possible and not overfitted. Overfitted models are
not able to predict variables due to their reflection of random
relationships between variables.
- Output of an analytical model as a explanation component. That
means, you can evaluate the analytical model to interpret the
obtained results immediately after modeling. You don't have to
guess why results are as they are.
GMDH module of NeuroShell2 tool developed by Ward Systems
Group, Inc. use partial polynomial optimization for network structure
construction. ModelQuest software from AbTech Corp. realise
so-called abductive polynomial GMDH networks. Unfortunately the
recent developments and modern GMDH experience were not yet realized
in the software considered above. They turn into account over-complicated
models using heuristics, which contradicts to self-organizing nature
of the approach. Often is used an internal criteria and therefore this software
products are successive in the case when dispersion of noise in
data is small or data sample is long. In opposite case more effective
are original inductive GMDH algorithms based on external criteria
Now the most devised software developed under this approach is
by Frank Lemke for Apple computers.
This modeling tool realizes twice-multilayered
neuronets with active neurons, optimizing the structure of every
neuron (COMBI) and adaptively
synthesizing the network structure (MIA).
It can be used for generation of systems of equations (OSA)
The AutoNet is Excel-based application with open VBA code from Peak Consulting Inc. This is GMDH-type neural network program with a self-designing
The main SKAT and Polynet Predictor module in the PolyAnalyst software from
Megaputer Intelligence use GMDH approach for knowledge discovery and data mining. Their
main SKAT module use inductive GMDH-type technique, with sorting
of ratio-polynomial models.
Fuzzy GMDH is used in DataX data mining software from Zaptron Systems, Inc.
GMDH for Java - artificial GMDH-like network written entirely
in Java is available from Andy Zelezny's homepage. It's not based
on inductive principles unfortunately.
Software which realize inductive approach is not complex. That's
why there are a lot of "raw" programs-prototypes, written by many
investigators throughout the world, which realize simplified multilayered
GMDH algorithms. Some of them are listed here.
We ask all who have such software to contact with us, your work
may be helpful for another researchers.
In Russia there were created three different commercial mainframe
packages of Applied Programs (PPP) in NPO "TsentrProgramSystem"
in Tver. It were widely used in many organizations mainly for complex statistical
analysis of time series using GMDH.
Rather good old implementation of multilayered GMDH algorithm Astrid
for DOS only is available from our Glushkov Cybernetic Center.
There is still necessity in further development of GMDH software
under Windows OS in which modern GMDH experience will be used. For
broadening of functional software possibilities the construction
principle of multi-functional algorithmic modules is proposed in
: generalized structure
generator, generalized model, generalized criterion. This modules
under supervision of coordinating monitor gives broad spectrum of
known methods of structural identification, regression construction
and restoration of dependencies.