Investigation of Relationship between Learning Amount and Interval Solution Range in Multi-Objective Optimal Design Using ANN Model

In order to efficiently perform a redundant design that can simultaneously satisfy conflicting performance requirements and cover a wider range of design parameters, we investigated the relationship between the amount of ANN training data and the magnitude and accuracy of the interval solution range in multi-objective optimal design using an artificial neural network (ANN) model, with the target being the design of an electromagnetic interference (EMI) filter for a brush motor drive system. In this study, it was found that when the design parameters are at 2 levels, the maximum and minimum values, the amount of training data is very small but a range solution is still obtained though it is insufficient. Increasing the amount of training data increases the training time for the ANN and the number of false positives (FPs). However, since the possibility of FP occurring in the range solution is low, increasing the amount of training data to obtain a wide range solution is effective. Furthermore, since the training time increase significantly when the amount of training data is increased, this study concludes that 3 to 5 levels are effective.