NN Program Set for Optimal Design of Material
Objects
S.I.Rodin
Department of Material Science and Mechanics
The methods of genetic (evolutionary)
algorithms (GA) are used for search of optimum solutions. The special
interest represents application of these methods to search of optimum
solutions on the basis of evolution directly to material objects of the
environmental world.
See also: Experiment
Forecast with NN and GA
At study of such
objects in any field of science and engineering the models are used, at
which hypotheses simplifying a structure and connection of the researched
phenomenon are introduced. It is supposed that the introduction of such
hypotheses does not take into account minor, insignificant for the given
task, properties of an object.
These assumptions are checked experimentally
and thus never it turns out of exact concurrence to theoretical results. The
presence of small deviation theory-experiment serves criterion of a
correctness of theory and basis for its application in practice. It is impossible to take into account
infinite connections of the environmental world and the given approach is
justified.
The direct application of evolutionary
methods to material objects doesn’t require models and hypotheses.
Thus can be detected and used new phenomena, unknown today, that can result
to qualitative new level of knowledge about the environmental world and
creation essentially new materials and products. A
key benefit of neural networks (NN) is that you can use them to build a
model of the system or subject you are interested from just the data you
provide them. You know the inputs and outputs that are important but may not
know what happens internally; well the neural network will model this system
for you from the data. Neural nets are powerful solutions to
these problems.
 |
NN Program Set consists of
4 programs:
 |
NNdesign – design and train NN on base of GA. |
|
NNoptimum – find optimal solution
for multiobjective problem on base of NN with GA. |
|
NNcalc – viewer and calculator
for NN. |
|
GAreal - find optimal solution
for multiobjective problem on base of micro-GA for a very small population of
material objects with experimental study for every generation - direct
measurements in the real world are used. |
|
All these programs use common Excel
book from MS Office for input data and output results, on which Worksheets
all data are presented.
On Download
Demo Page you may download NN Program Set demo and see how programs work. There are
instructions how to install.
Program NNdesign
This program is used for design and train NN
with GA. Island ring topology is used for optimization. During evolutionary
process optimal NN is searched: structure and node parameters. On Fig.2 examples of NN are shown.
|
Full functional copy of program NNdesign:
|
uses Excel from MS Office for input
data and output results; |
|
NN structure may have any number of
nodes; |
|
NN may have any number of Input and
Output nodes; |
|
may have any number of training data
series; |
|
uses flexible transfer functions for
nodes – X^n, EXP, LN, SIN, 1/X – user may form any combination; |
|
stores in Excel file results of
training and optimization – real output data series, structure
ties, weights for node ties and transfer function parameters for
nodes. |
|
In common case transfer function for nodes is
where  - table function
(Fig.1),
values are randomly generated for initial set with any given number of
arguments, its parameters are determined during evolution and are different
for different nodes;  - analytical
function.
Fig. 1. Table function
All parameters of transfer function are
determined during evolution and are different for different nodes; user may
exclude any part of it.
The use of table function makes possible to
find and use the best transfer function for the given task.
 
 
 
Fig. 2. Neural networks
Program NNoptimum
This program is
used to find optimal solution for multiobjective problem on base of NN with
GA for island ring topology. NN is the result of program NNdesign.
|
Full functional copy of program
NNoptimum:
|
uses Excel from MS Office for input
data, restrictions, NN parameters and output results; |
|
user may introduce restrictions for
min/max values of input and output data; |
|
user may exclude several inputs and
outputs from optimization process setting appropriate constant
values for such inputs with possible restrictions for such outputs; |
|
user may define every output so as to
find min or max for it; |
|
stores in Excel file results of
optimization for multiobjective problem. |
|
Program NNcalc
This program is used to view and save
into graphic file structure of NN produced by program NNdesign and to
calculate NN for any input values to analyze results.
Program GAreal
Program is used to find optimal solution for
multiobjective problem on base of micro-GA for a very small population of
material objects with experimental study for every generation.
No any models or hypothesis for members of
population are involved. Direct measurements in the real world are used.
It’s possible to use help of NN for every generation to reduce number of
real world measurements.
Program uses Excel from MS Office for
input data and output results.
Fig. 3. Optimization process
|
The procedure of optimization may be
following (without GAreal):
|
Build NN with NNdesign program
and initial training data series.
|
|
Find optimal solution for
multiobjective problem with NNoptimum program for obtained
NN.
|
|
Make laboratory measurements for
optimal data set. Sure, output data would differ from optimal solution based on NN.
|
|
Add new data (last optimal input and
laboratory measurements output) to initial training data series and
build new NN. |
|
Program GAreal gives possibility to
avoid any models and hypothesis: just look, measure and go to the next
generation. Above optimization process with NN may be added for help to
reduce number of real world measurements (Fig.3).
This procedure and NN Program Set were
used to search for composition and production technological parameters of
road surface material for freezing temperatures. It was multiobjective
problem with several constraints.
Samples production and laboratory
measurements took several days for one generation and the whole study took several
months.
The obtained results were unpredictable;
properties and quality of new material were much higher of existing ones.
Additional research showed that it was due to special microstructure for
obtained material.
See also: Experiment
Forecast with NN and GA
On Download
Demo Page you may download NN Program Set demo.
Details by E-mail.
|
