UDC 338.24

DATA PROCESSING BY NEURAL NETWORKS

Imanbayev K.S., Adil A.T.

ATLMATY TECHNOLOGY UNIVERSITY

The article deals with hidden, non-trivial and non-formalized regularities in data sets, where there are large information arrays of heterogeneous, time-varying data. With the help of the neural network model, the parameters of the model, as well as the parameters of training the neural network, have been preprocessed.

Keywords: forecasting, neural networks, hidden, nontrivial and non-formalized regularities, exponential smoothing.

Identifying hidden, non-trivial and non-formalized regularities in data sets, where there are large information arrays of heterogeneous, time-varying data of high dimension is a difficult task for decision-making. As you know, there are five types of patterns: association, sequence, prediction, classification, clustering. In this paper, we consider the prediction problem. The task of forecasting is based not only on the previous values of the predicted value, but also takes into account the influence of various additional factors represented by time series. The task of forecasting has the following features:

- the predicted value is affected by several different factors

(Also being time series) in which the analytical.

The description of these dependencies is difficult;

- the time series in question are nonlinear and

Nonstationary:

- the form of the nonlinearity of the time series is a priori unknown and not described analytically;

- as a result of solving the problem, it is required to find a non-smoothed short-term forecast.

Traditionally, the methods of mathematical statistics are used to solve similar problems.

Exponential smoothing is a popular method of predicting time series has the form:

X = b + εt

B - constant, ε - random error slowly changing with time. The exact formula for simple exponential smoothing has the form:

St = (αXt + 1-α) S t-1

St - smoothed value, Xt - current value, α-parameter

When this formula is applied recursively, each new smoothed value (which is also a forecast) is calculated as the weighted average of the current observation and the smoothed series. Empirical studies have shown that very often simple exponential smoothing gives a fairly accurate prediction. The autoregressive integrated moving average is based on the combined use of two time series models:

- autoregression process;

- moving average process.

The autoregression process contains elements that depend on Friend. This dependence is expressed by the following equation:

X = ξ + φ1 X_t-1 + φ2 X_t-1 + ... + ε

where ξ is a constant (free term), φ1, φ2 are autoregression parameters.

Each observation is the sum of a random component (random Impact) and a linear combination of previous observations.

The moving average process represents the case where each.

The element of the series is subject to the combined effect of previous errors. AT

General form can be written as follows:

Xt = u + εt-θ₁ ε_t-1 - θ ₂ ε_t-2

Where u is a constant, θ1, θ 2 - parameters of the moving average, ε - error

The current observation of the series is the sum of the random component (random action of ε) at a given moment.Multiple linear regression, like a straight line in the plane, looks like this:

Y = a + b * X

The regression coefficients a, b represent the independent contributions of each independent variable in the prediction of the dependent variable. The regression line is constructed in such a way as to minimize the squares of deviations of this line from the observed points and to obtain the equation of a straight line that best describes the dependence of Y on X.

Graphical technical analysis is used to determine the probability of continuation of graphic models:

1) Models of the fracture of the trend - the models formed on the graphs,

Which, under certain conditions, may prejudge a shift

Existing on the trend market.

2) Models of continuation of the trend - formed on the graphs

The models, which, under certain conditions,

That there is a high probability of continued

The current trend.

A major drawback of graphic technical analysis is that it is very subjective.

The features described above limit the possibilities of application

Statistical methods, as well as the application of various smoothing procedures. Does not correspond to the task in view, since in forecasting we are not interested in the smoothed value, namely the deviation from it at a future moment in time. The predicted time series is characterized by the fact that it is not stationary and does not transform to a stationary one. Designed for stationary series, for non-stationary series, they can not be used. Linear regression models are limited in that they do not always take into account the nonlinearity of the process. Technical analysis tools are a common way of predicting time series. The advantage of which is ease of use and visibility. In total, there are now more than 160 different indicators and methods. Technical analysis is in the properties of flexibility and adaptability, as well as the strength of the possibility of its application at any point in time. In general, technical analysis is characterized by subjectivity of experts (especially in graphical analysis), the influence of false signals, the inconsistency of various indicators.Technical analysis is based on the predicted series, without explicitly using information on other factors. The methods of mathematical statistics and technical analysis traditionally used for forecasting give good results in many cases, but neural networks have several advantages:

- the use of neural networks does not impose any restrictions on the nature of the series under study, the nonstationarity of the processes under consideration does not present a problem. Success of solved problems, in which it is difficult or impossible to find analytical dependencies between input and output data. Finding the optimal indicators for this task and building on them an optimal prediction strategy for the given series.

- The strategy can be adaptive, changing with the situation.

It is mathematically provable that they can represent any real continuous function of any real continuous vector argument. Multilayer networks can be used to solve any

Task, which can be reduced to the construction of functions, including

For forecasting.

Neural networks have the following disadvantages:

  - the need to attract high-quality specialists, in view of the high

The complexity of this setting, the lack of guarantees for a successful solutionTask.

- existing software does not have a specific problem.

Orientation and are not adaptive to the solution of the class of problems under consideration. To select the type of neural network used in this work, an analysis is made of the applicability of specific types of neural networks to the solution of various classes of problems. The result of this analysis was the definition of neural network architectures that can be used to solve the problem of predicting a multilayer perceptron. The forecasting technique used by this work is the main one on the account of the unformalized dependencies between various factors in the array of empirical information. Thus, without imposing any restrictions on their character. To improve the performance of the neural network model, preliminary processing of data, determination of model parameters, as well as training parameters were carried out. The quality of the model obtained, the learning algorithm taking into account the specific nature of the problem being solved is estimated. Preliminary preparation of data is the basis for the successful solution of the problem, for which a method of such preparation is proposed. After that, the information content of the initial data is increased. The following is suggested:

-transformation of data for the sensitivity of the network;

  -application of an effective method of preliminary data processing to enhance the neural network model;

- definition of model parameters;

- determining the parameters of training;

- evaluation of the quality of the obtained model for constructing and determining the structure.Forecasting system.

Bibliography:

1. Robert Callan. Basic concepts of neural networks [Text] -

"Williams", 2001. - 288 p.

2. Eremin D.M, Garceev I.B. Artificial neural networks in

Intellectual management systems [Text] - M .: MIREA, 2004.

- 75 seconds.

3. Simon Khaikin. Neural networks: a full course [Text] -

M .: Williams, 2006. - 1104 p.

4. Berkinblit M. B .. Neural networks [Text] - M .: MIROS and VZMSh

RAO, 1993. - 96 p.

5. V.A. Terekhov, D.V. Efimov, I. Yu. Tyukin. Neural network systems

Management [Text] - Higher School, 2002. - 184 p.