ИСТИНА |
Войти в систему Регистрация |
|
ИПМех РАН |
||
Nearly all experimental results in modern science are the results of indirect measurements. This means that there is a separate problem called an inverse problem (IP) - to extract the information interesting for the researcher from the measured experimental data. Solving such problems is an inherent necessity in many areas of science, including spectroscopy, geological prospecting, aerospace image processing and others. The lecture discusses methodological aspects of the solution of IP with the help of such biologically inspired cognitive architecture as artificial neural networks (ANN). Different formulations of IP from the point of view of data processing methods are given. Various methodological approaches to the solution of IP using ANN techniques called “experiment-based”, “model-based”, and “quasi-model” approaches, are considered. Their characteristics, differences and areas of application are discussed. The differences of ANN from other methods of solution of IP are discussed, as well as the key areas where their use is justified. Different approaches to simultaneous determination of parameters when solving multi-parameter IP are considered. The material is illustrated by examples of IP from the areas of optical spectroscopy and electrical prospecting.