Title: Approximation and Prediction of Wages Based on Granular Neural Network
1Approximation and Prediction of Wages Based on
Granular Neural Network
- Milan Marcek1 and Duan Marcek2
- 1Faculty of Philosophy and Science, Silesian
University, 746 01 0pava, Czech Republic MEDIS
Nitra, Ltd., Pri Dobrotke 659/81, 949 01
Nitra-Draovce, Slovak Republic - marcek_at_fria.utc.sk
- 2Faculty of Philosophy and Science, Silesian
University, 746 01 0pava, Czech Republic
Faculty of Management Science and Informatics,
University of Zilina 010 26 Zilina, Slovak
Republic - dusan.marcek_at_fpf.slu.cz, dusan.marcek_at_fri.uniza.sk
2Outlines
- Basic principles of identifying input-output
functions of systems and forecasting - RBF and soft RBF NNs for approximation of input -
output functions - Additive fuzzy system
- Granular RBF network based on cloud concept
- An application (experimenting with statistical
and eonometric models vs. RBF networks) - Results
- Conclusions
3Basic principles of identifying input-output
functions of systems and forecasting
There are two major approaches to forecasting
explanatory and time series.
Explanatory forecasting
Time series forecasting
(ANN)
4RBF and soft RBF NNs for approximation of input -
output functions and forecasting
a)
b)
5Additive fuzzy system
The fuzzy system consists of series of separate
fuzzy rules (relations) each of the type of if
Ai then Bi. Centroidal output converts fuzzy
sets vector B to a scalar. The most popular
centroidal defuzzification technique uses all the
information in the fuzzy distribution B to
compute the crisp y value as the centroid or
centre of mass of B where y stands for the centre
of gravidity of the jth output singleton.
6Granular RBF network based on cloud concept
Cloud models are described by three numerical
characteristics Expectation (Ex) as most
typical sample which represents a qualitative
concept, Entropy (En) as the uncertainty
measurement of the qualitative concept and Hyper
Entropy (He) which represents the uncertain
degree of entropy. En and He represent the
granularity of the concept, because both the En
and He not only represent fuzziness of the
concept, but also randomness and their relations.
Then, in the case of soft RBF network, the
Gaussian membership function
7An application (experimenting with statistical
and eonometric models vs. RBF networks)
a)
b)
B-J methodology ARMA(1, 3) process yt
-0.0016557 -0.4567y(t-1) et 0.90516e(t-1)
0.58768e(t-2) 0.36497e(t-3)
MSEA
0.014, MSEE 0.048
Ekonometric model (transfer functions model) yt
0.239347 1.04044 yt-4 MSEA 0.0026, MSEE
0.0033 (7)
8Results
9Conclusions
- The estimated parameters, in contrast with
statistical models, have no economic
interpretation to structural model, all
parameters in the model are fixed, and there is
no possibility to test the stability of the
parameters - The econometric model in (7) gives best
predictions outside the estimation period and
clearly dominates the other models. We have shown
that too many model parameters results in
overfitting, i.e. a curve fitted with too many
parameters follows all the small fluctuations,
but is poor for generalisation
10Algoritm for updating weights in the granular
neural network
11