Over-fitting of noisy teacher data is a serious problem for artificial neural networks (NNs). Their learning method based on Bayesian inference was proposed by MacKay in order to prevent over-fitting and generate a smooth response surface (RS). Using this method, the response surface of neural networks can be used to predict optimal input conditions, because it has good generalization (interpolation) ability. We specified and verified the learning algorithm and the modification method of hyper parameters, which are coefficients that appear in an objective function with penalty terms. After three different modification methods were applied to interpolation examples, we chose the superior modification method. Finally, the response surface of neural networks using the superior method was applied to an optimization problem in which the fiber orientations of a laminated composite structure were determined so as to minimize the difference in two displacements. We concluded that the learning algorithm and the modification method are effective to optimization problem.
抄録全体を表示