|
In my master's thesis the problem of forecasting with the help of neural networks is investigated. A practical part of the given work is creation of the program – realizations of a neural network, and learning of the given neural network. The data for learning is time series or values of any function.
Recently it is observed interest to use of neural networks for the decision of various problems and their use in different spheres of a human life. With use of neural networks opportunities of use of calculations in spheres, before concerning only to area of human intelligence, an opportunity of creation of computers which ability to study and remember wonderfully reminds thought processes of the human have opened.
Artificial neural network – a set of neuron, connected among themselves. Artificial neuron on the properties reminds biological neuron.
On an input of artificial neuron come in some set of signals x1, x2, …, xn, each of which is an output of another neuron. Each input is multiplied on the corresponding weight w1, w2, …, wn, and come in on the summing block Σ. The summing block corresponding to a body of a biological element, adds the weighed inputs algebraically, creating an output – NET.
The elementary neural network calls perseptron. On the structure it is similar to neuron, but on its outputs there is some analyzer. And depending on value of the generated sum, target value of the analyzer will be equal "1", if the sum more than threshold value and "0" if it is less.
Today perseptron is one of the most popular realizations of neural networks. The reason of its popularity is relative simplicity of realization on a background of universality and the broad audience of problems which can solve perseptrons.
For the solvation more difficult problems use multilayered neural networks. But for training multilayered neural networks required more difficult learning algorithms.
Learning of the artificial neural network is some process which modify its weight. If learning is successful, giving to a network set of input signals brings in result desirable set of target signals. There are two classes of learning methods: necessitarian and stochastic.
The necessitarian method of learning step by step carries out procedure of correction of weights of the network, based on use of their current values, and also weights of inputs, actual outputs and desirable outputs.
The stochastic method of learning carry out pseudo-casual changes of weights, keeping those changes which conduct to improvements.
Forecasting is a prediction of the future events. The purpose of forecasting is reduction of risk at decision-making. In most cases the forecast turns out erroneous, and the error depends on predicting system and methods of forecasting. For reduction of a mistake it is necessary to increase quantity of resources given for the forecast. At some level of a error it is possible to achieve a minimum level of resources for the forecast. The basic problem of forecasting is revealing discrepancy of the forecast. Usually, the decision accepted on the basis of the forecast should take into account a error on which the system of forecasting informs. Thus, the system of forecasting should provide definition of the forecast and a error of forecasting.
The most widespread problems of forecasting, for today are problems forecasting in economy and at the enterprises, and in particular financial planning, planning of technological process, planning of the share market, etc.
Forecasting with neural networks has a numberof lacks. As a rule it is necessary about 100 supervision for creation of comprehensible model. This big enough number of the data, also exists many cases when such quantity of the historical data is inaccessible.
However, it is necessary to note, that construction of satisfactory model on neural networks even in conditions of shortage of the data is possible. The model can be specified as the new data becomes accessible.
|
