An
I-section steel welded girder with four vertical ribs is considered. Three
amplitudes of smooth shape
imperfections
of the web plate are assumed as random
variables. A hybrid
approach is used in which trials in the Classical Monte
Carlo method (CMC) are simulated by a trained Back-Propagation Neural Network
(BPNN). Sets of patterns for BPNN training and testing are computed by the
COCMOS/M nonlinear Finite Element Method (FEM) code. It is shown that a great
increase of numerical efficiency can be obtained due to application of BPNN in
the hybrid approach CMC-FEM/BPNN.