D.I.Plienkina, S.L.Nikulin
National mining university,
Dnepropetrovsk, Ukraine
Efficiency
of use of different recognition
methods at the decision of geological problems
Nowadays
the methods of pattern recognition are widely applied to the decision of geological
problems, but efficiency of their use is not studied in details. Thus the
problem of estimation of recognition quality attracts not less attention, than
tasks, related directly to recognition. In practice concept "quality of
recognition" usually expresses the relation of the expert to result and is
defined by degree of conformity of classification to the fact data. As indices
of such accordance we can use absolute or relative error of result, the
confidential probability of belong decision to the set area, degree of
coincidence of prognosis estimations of some parameters with true ones, etc. In
any case the question is about similarity of decision to true one or about the
error of prognosis (recognitions).
The
mathematical model of prognosis is usually described as 
, where 
– is a result of prognosis; 
– is a set of prognosis methods (decision rules,
operators, algorithms); X – is initial data; 
– is an error of
prognosis.
The
method of quality prognosis estimation substantially depends on its purpose,
appointment, a priori information and accepted assumptions.
The purpose of the present
work is determination of efficiency of use of different recognition methods by
the procedures of quality estimation. As the decided task the problem of
prognosis of deposits at Ziaetdin ore-gold (Uzbekistan).
Recognition and estimation of quality was carried out in the specialized
geoinformation system RAPID [2],
developed at Geoinformation systems department at the National mining university (Dnepropetrovsk, Ukraine).
In the RAPID system the problem of recognition of geological data decides by followings methods:
– logical ("Kora-3" algorithm);
– statistical (a parametrical method of estimation of
distribution density, a nonparametric method of estimation of distribution
density);
– deterministic (similarity measure function,
potential function, angle measure function);
– frequency (recognition on the basis of relative frequency,
recognition on the basis of calculation of forecasting function);
– neural networks (Rozenblatt perceptron, multilayer
perceptron, radial-basis neural network, probabilistic neural network, support
vector machine, Elman recurrent network, second-order recurrent network).
In this
work the quality estimation is considered as the process consisting of three
consecutive stages:
1.
Defining the purpose of the estimation.
2.
Setting prognosis attributes desirable for achieving recognition results, or defining quality criteria.
3.
Forming quality measures and estimation rules used for quantitative evaluation
of desirable prognosis results.
At
recognition a prognosis is usually produced in a categorical form. The most
widespread criterion of quality in this case is number of the errors, which are
counted up for objects of training or control sample.
In the RAPID system following indicators of recognition quality are implemented: type I errors, type II errors, risk of searches, ratio Brayer’s
index, logarithmic index, spherical index.
The
listed sizes allow to receive quantitative information about the results of
prognosis of objects of certain class, and also the average values for all
classes.
By
results of the carried out research the nonparametric method of an estimation
of distribution density has shown the greatest efficiency. The error of
prognosis was 5,64 %.
Literature:
1. Busygin B.S.,
Miroshnichenko L.V. Recognition of patterns at geologo-geophysical
prognostication. – Dnepropetrovsk: Publishing house DSU, 1991. – 168 p.
2. Pivnyak G.G.,
Busygin B.S., Nikulin S.L. GIS-technology of the integrated analysis diverse
and multilevel geodata. // Reports of National academy of sciences of Ukraine,
2007. – ¹6 – p. 121-128.