Neural-network assistance to calculate precise eigenvalue for fitness evaluation of real product design

When applying genetic algorithms (GAs) in the design real products, reducing the computational cost of fitness functions is one of the major challenges. In some cases, the computational cost of calculating specific eigenvalues is a predominant factor and needs to be reduced. We proposed the use of a GA with "neural-network (NN) assistance," which enables this computational cost to be reduced. With this GA, the NN assistance infers the approximate eigenvalues. Then, these approximate eigenvalues are used when starting the convergence calculation to obtain the precise eigenvalues. This procedure is effective in reducing the total computational cost of some of the fitness functions of real products. In addition, the precision of the eigenvalue is retained because the precise eigenvalues are obtained by the convergence calculation. The results of our case study show that the GA using our method achieves a 3x speed-up in fitness computation while maintaining equivalent solution quality.