Scheme of establishing of optimal structure of supervized-trained multilayer perceptron is proposed basing on computation and minimization of Bartlett’ or NIC generalization estimators during explorational search in the space of elementary structure-modifying operations, the best of which is then applied at the current epoch. Number of structure-modifying operations is extended comparing with well-known constructive algorithms. Experiments are conducted for four real-world supervised classification tasks.
Note. Abstracts are published in author's edition
Webmaster: www@www-sbras.nsc.ru |
|