Experiment Forecast with NN and GA
S.I.Rodin
Department of Material Science and Mechanics
Experimental results depend on many factors.
Some of them are known and may be taken into consideration by some methods.
The influence of weak known such as daytime, solar activity, moon faces and
many others is difficult to understand. And it’s impossible to understand
the influence of unknown infinite connections of the Universe.
Here is introduced method for prediction of
experimental results without models and hypotheses on the basis of two
databases: “past” and “present”. These may be the results of the
same phenomenon on different time steps (last year or month and now, for
example) or it may be measurements not connected with time for similar
objects. Such objects and phenomena may be of any nature: mechanical
properties of construction materials, chemical reactions, water supply and
wastewater, stock exchange, sport competitions, games etc.
On
the basis of these two databases is designed model to transform “past”
to “present” and to forecast future data. Neural networks (NN) are used
for this purpose designed with genetic algorithms (GA). See details: Neural
Networks and Genetic Algorithms in Optimization Problems.
Elements of database “past” are input
data for design and training NN (equals to N) and initial data for forecast
(equals to M). Elements of database “present” are output data for design
and training NN (Fig.1). Number of inputs of NN is equal to the
number of measurements for every element database “past” (Data 0-i).
Number of outputs of NN is equal to the number of measurements for every
element database “present” (Data 1-i). Number of inputs and outputs may
be different: time intervals or number of measurements in databases
“past” and “present” are various.
Fig.1. Databases “past” and “present”
Program NNforecast is used for design,
train NN with GA and forecast calculations (is based on program NNdesign).
Island ring topology is used for optimization. During evolutionary process
optimal NN is searched: structure and node parameters,
 | Full Program NNforecast has
following features:
 |
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, including table function; |
|
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, forecast. |
|
Additional program NNcalc is used to
view and save into graphic file structure of NN produced by program NNforecast
and to calculate NN for any input values.
In the Table 1 and on Fig.2
(initials and forecast – solid curves) are shown results of 24
measurements for elements of every database and forecast (N=3, M=1). On this
base is designed NN (Fig.3) and calculated forecast.
Table
1
| action |
training
IN data |
initial |
training
OUT data |
forecast |
| day |
0-1 |
0-2 |
0-3 |
0-4 |
1-1 |
1-2 |
1-3 |
1-4 |
| time |
past |
present |
| 1 |
51,92 |
48,19 |
48,93 |
49,68 |
38,94 |
37,36 |
34,46 |
37,24 |
| 2 |
45,73 |
47,94 |
50,58 |
48,08 |
34,29 |
37,19 |
32,80 |
35,16 |
| 3 |
45,43 |
43,63 |
45,04 |
44,70 |
34,07 |
29,39 |
23,77 |
29,12 |
| 4 |
44,68 |
45,99 |
49,51 |
46,73 |
33,51 |
28,74 |
25,37 |
28,37 |
| 5 |
41,98 |
44,64 |
42,31 |
42,98 |
31,48 |
32,80 |
28,90 |
31,03 |
| 6 |
41,53 |
39,55 |
43,00 |
41,36 |
31,15 |
28,13 |
30,49 |
29,91 |
| 7 |
41,18 |
42,73 |
44,55 |
42,82 |
30,88 |
30,39 |
26,69 |
30,70 |
| 8 |
43,08 |
41,85 |
41,43 |
42,12 |
32,31 |
32,02 |
29,08 |
30,54 |
| 9 |
48,42 |
51,47 |
51,48 |
50,46 |
36,32 |
34,76 |
40,18 |
37,80 |
| 10 |
49,07 |
47,61 |
49,52 |
48,73 |
36,80 |
32,54 |
32,57 |
33,87 |
| 11 |
49,47 |
47,23 |
44,15 |
46,95 |
37,10 |
35,50 |
31,84 |
34,43 |
| 12 |
56,57 |
56,00 |
56,29 |
56,29 |
42,42 |
38,10 |
38,76 |
40,14 |
| 13 |
47,22 |
49,56 |
51,98 |
49,59 |
35,42 |
26,81 |
28,16 |
28,97 |
| 14 |
46,72 |
47,04 |
49,21 |
47,66 |
35,04 |
34,27 |
32,21 |
33,77 |
| 15 |
51,12 |
49,57 |
45,84 |
48,84 |
38,34 |
36,00 |
41,18 |
38,43 |
| 16 |
49,97 |
49,11 |
50,97 |
50,02 |
37,48 |
34,84 |
30,68 |
33,57 |
| 17 |
56,42 |
52,94 |
55,67 |
55,01 |
42,31 |
37,26 |
37,63 |
39,77 |
| 18 |
55,72 |
55,51 |
56,61 |
55,94 |
41,79 |
37,29 |
37,85 |
38,95 |
| 19 |
56,52 |
59,49 |
61,70 |
59,23 |
42,39 |
36,79 |
34,73 |
36,78 |
| 20 |
55,67 |
58,10 |
54,58 |
56,12 |
41,75 |
38,68 |
43,64 |
41,51 |
| 21 |
56,07 |
53,92 |
54,99 |
54,99 |
42,05 |
39,68 |
34,93 |
39,04 |
| 22 |
56,02 |
55,33 |
52,60 |
54,65 |
42,01 |
37,46 |
40,20 |
39,88 |
| 23 |
63,26 |
65,81 |
68,71 |
65,93 |
47,45 |
44,45 |
43,18 |
45,66 |
| 24 |
60,00 |
62,56 |
59,81 |
60,79 |
45,00 |
40,72 |
43,19 |
43,00 |
 
Fig.2. Databases “past” (left) and
“present” (right),
initials and forecast – solid curves
Fig.3. NN for Table 1
Such forecast method may be applied to wide
range of objects and phenomena.
On Download
Demo Page you may download Help_forecast zip file with
help file and Excel file with initials and results.
Details by E-mail.
|
|
