The goal of the development is a making the site "Assessment of real estate on neural network" for granting user possibility of the study of the procedures of the assessment of real estate and analysis of the assessment of real estate on neural network. This site gives a user a possibility of study of the procedures of the assessment of real estate: spent method, a revenue method, a method of the benchmark analysis of the sale. Also undertaking the analysis of the assessment with use trained neural network.
The core of this paper is analysis of assessment of real estate in Volgograd at the end of year 2006. This analysis based on artificial neural nets modeling. We was investigated predictive index of real estate requirements. Estimation was carried out using Deductor (neural modeling tool). Effective model of neural nets for forecasting problem solving is multi-layer perceptron. During learning neural net with back-propagation algorithm, neural net is capable to made most probable forecast if learning sample is well exact and wide. Base for learning sample was created from data of Federal Service of Staticstics from 1995 till 2004. As an input information were used following activities: average coefficient of flats´ price increasing (%); average price of rent of holding (RUB per 100 square meters); average costs for forming housing (thousands RUB per square meter); commission of blocks of flats (million square meters per year). As in output information was used index of blocks of flats need per year. As units of blocks of flats need we take minimal space of housing (12 square meters). Macros BG_ExportToDeductor.xla allowed to export data to Neural Analyzer 3.0 (one of Deductor´s modules). There are statistical sampling data are tuned in this window. Field 1 is informational because it contain information about years of statistical examina-tion. There are net parameters settings in this window: algorithm choice, function choice, steepness choice, and so on. In this case back-propagation algorithm was choose. It is necessary to make up learning sample for experiment. In this window statistical data are inputed for operative experiment with learned neural net. Neural net is consider learned if results of experiment agree closely with testing samples. If it isn´t true then it is necessary to re-learn neural net.
Misprediction is from 4.29% till 10.34%. There are following data receiving from analysis second half year 2006. Average price of rent of holding 261%; average costs for forming housing (thousands RUB.) - 15180; commission of blocks of flats (million square meters per year) - 39; coefficient of price changing 119, 8 %. Judge by neural net´s forecast, commission of blocks of flats will be 4598. As neural net was mistaked, a result will from 4001 till 4598.
The work is submitted to the International scientific conference "Modern science technology", Tenerife, Spain, November, 20-27, 2008, came to the editorial office on 01.06.2008.