| DonNTU > Masters' portal > Igor Yemeliyanenko | RUS | ENG |
| Main page | Autobiography | Library (RUS) | Links (RUS) | Search report (RUS) | Individual task (RUS) |
The degree work was not completed at the time of writing the author's abstract . January 2007 is a time of the completion of writing the work. A full-text version of the degree work and other materials on its subject can be received both at author and head.
The forecasting is a key moment in making decisions management. A final efficiency of any decision depends on the sequences of events which appear after decision has been already made. The possibility to predict uncontrolled aspects of these events before decision making allows to make the best choice, which, otherwise, could be not such ingenious. That is why planning system and management usually realize the function of forecast.
The explosion of interest to neural nets exists in the last several years, they find successful using in the most different areas - in business, medicine, technology, geology, physics, etc. Neural nets enters into the practice everywhere, where it is necessary to solve problems of the forecasting, categorizations or management as far as they applicable practically in any situation when there is relationship between variables-predictors (inputs) and forecasted variable (outputs), even though this relationship has very complex nature and it is difficult to express it in usual terms of correlation or differences between groups.
The typical example of neural net with a baseband of the signal is shown on the drawing below. Neurons are organized in layers by regular image. The input layer serves for entering of input variables meanings. Each of hidden and output neurons united with all elements of a previous layer.
Networks with complete system of relationships are preferred for majority of applications (such relationships are used in software package STATISTICA Neural Networks).
While neural net is working, meanings of input variables enters input elements, then neurons of intermediate and output layers work off consecutively. Each of neurons forms its output after processing of signals. Output meanings of element of output layer are taken for the output of the whole network after the whole network will work off.
Traditional statistical models are important class of the models, which are offered to a researcher by mathematics. By means of these models are described phenomenas, in which statistic factors, that do not allow to explain the phenomena in purely deterministic terms, are present.
Neuronal nets (NN) can be expressed in two ways: the first is a program model of NN, the second is a hardware. Products, which are founded on using of the mechanism of NN's action, originally came up for type of microcircuits.
Neuro-BIS are and, probably, will stay the basic commercial hardware product on base of NN in the near future.
Neuro-BIS will, certainly, become the base of a new neurocomputers and specialized multiprocessor products. The majority of present-day neurocomputers represent itself as simple personal computers or workstations in composition with an additional neuronal microcircuit.
NN well suit for artificial perception and decisions of problems to categorizations, optimization and forecastings.The list of the possible industrial usings of neural nets, on the base of which are already created commercial products or are made demonstration prototypes, is brought below:
- banks and insurance companies;
- administrative service;
- petrolium and chemical industry;
- military industry and aeronautics;
- industrial production;
- security service;
- biomedical industry;
- TV and communication.
The presented list is far from being full.
Problems of my work are:
- execution of forecasting of meanings of some social-economic factors on alike raw datas with use both of traditional statistic and neural nets' methods;
- execution of comparative analysis of forecasting results' efficiency on determined criterias over each of selected methods;
- determination of that, which type of the problems suits each of selected methods for forecasting best of all.
Executing of a complex comparative analysis of statistic and neural nets' methods for forecasting is a novelty of my degree work .
Methods for forecasting are possible to divide into two classes (kvalitative and kvantitative) depending on that, what mathematical methods are used.
Kvalitative procedures produce a subjective estimation which is founded on experts' opinion.
Kvantitative procedures for forecasting declare obviously how the forecast is received. Methods pertained to kvantitative procedure for forecasting are founded on statistical analysis, analysis of time sequences, Bayes's forecasting, set of methods, neural nets.
Two basic types of models are used now: time sequences models and causal models.
The time sequence is a ranked in time sequence of the observations of a variable.
The causal models use the relationship between the time sequence and one or more other time sequences.
Practically, forecasting systems often use the combination of kvantitative and kvalitative methods.
Following factors influence on selection of the method for forecasting:
- required form for the forecast;
- horizon, period and interval for the forecasting;
- accessibility of inputs;
- required precision;
- behaviour of the forecasted process;
- cost of the development, installation and work with system;
- simplicity of the work with system;
- managers' understanding and cooperation.
The row of measurements, which can be used for estimation of efficiency of forecasting systems, exists.
Amongst them the most important are: accuracy precision of the forecasting, system cost, resulting profit, characteristics of stability.
Simulation is an useful tool at estimation of different methods for forecasting.
The forecasting system must execute two main functions: generation of the forecast and the forecast management. Generation of the forecast includes:
- reception of inputs for the revision of forecasting models;
- undertaking the forecasting;
- the account of the experts' opinion;
- granting of results of the forecast to user.
Forecast management includes:
- observation of the process of the forecasting for determination of uncontrolled conditions;
- searching for the possibility for improvement of the forecast productivity.
Testing of a travel signalI is important component of management function. This function of management for forecast also must seasonly define the productivity of the forecasting and allow results to appropriate manager.
My own development at the completion of writing the degree work will touch the formation of an efficient method of making a comparative analysis of described in the work methods for forecasting over certain types of problems.
At completion of writing the degree work my experimental researches will contain:
- a real raw data for forecasting of meanings of certain gauges;
- executed forecasting of meanings of some social-economic gauges using several traditional statistic methods and by means of neural nets in such program packages as STATISTICA and STATISTICA Neural Networks;
- a comparative analysis (over determined criterias) of forecasting results of chosen traditional statistic and neural nets' methods.
On the grounds of foregoing it is possible to say that forecasting is a prediction of future events. The purpose of the forecasting is a reduction of the risk at decision making. The forecast is usually wrong, but error depends on used forecasting system.
The forecasting on NN possesses a number of lacks, such as a significant expenseses on time and other resources for building of satisfactory model.
However, in spite of enumerated lacks, neuronal models possesses a number of values, too. A suitable way to modify the model on measure of appearing of new observations exists. The model works well with time sequences. For this reason model can be used in areas, where we are interested in hourly, daily or weekly observations.
It is necessary to note that forecasting is not a final objective. The forecasting system is a part of a large management's system and, as a subsystem, it interacts with other components of the system playing considerable role in received results.
Perspectives of the further researches are rather encouraging, because mankind needs for exact mechanism for prediction of future more and more. That is why, it wll be so necessary to compare constantly and to select the best method of forecasting for the concrete problem or activity spheres. Over and above, neural nets become more complete, but together with them and neural nets' methods for forecasting.
The degree work was not completed at the time of writing the author's abstract . January 2007 is a time of the completion of writing the work. A full-text version of the degree work and other materials on its subject can be received both at author and head.
| Main page | Autobiography | Library (RUS) | Links (RUS) | Search report (RUS) | Individual task (RUS) |