АННОТАЦИЯ
Under consideration is further development of factor model analysis intended for monitoring of factors responsible for behavior of technical and other systems. Presented is a new technique of goodness-of-fit measure estimation for the case of unrestricted factor models employing arbitrarily distributed observed data, which is based on the capabilities of self-organizing feature maps (Kohonen networks) and the Monte Carlo method. Obtained results make it possible to avoid undesirable restrictions on observation data inherent in the traditional factor model identification procedure as well as restrictions on model structure used in the previous studies. Following this technique, each observed variance and covariance is associated with an equation that expresses analytically their expected value via free model parameters and equates it with the corresponding sample estimation. The overdetermined set of linear or non-linear equations is usually obtained and, then, solved with the aid of a numerical optimization procedure. Calculation of goodness-of-fit measure is based on comparison of the pseudosolution residual vector and generated random samples of residual vectors corresponding to the solutions being within the pseudosolution neighborhood. Simulated random samples of residual vectors are used to train self-organizing feature maps of proper dimension and, as a result, to obtain samples of Euclidean distances between residual vectors used as input cases and the centers (weight vectors) of "winning" units. That yields the opportunity to calculate the probability of exceeding the distance between the pseudosolution residual vector and its corresponding “winning” unit center and use it as a goodness-of-fit measure. A new measure for obtained pseudosolution precision as well as the technique for revealing the most probable factor model structure are under consideration.
ЦИТАТА
Kuravsky, L.S. Estimation of goodness-of-fit measures for identification of unrestricted factor models employing arbitrarily distributed observed data / L.S. Kuravsky, P.A. Marmalyuk, A.S. Panfilova // . – 2011. – P.
