Comparison of different approximation techniques. Complex problem with 25 design variables, only 1.000 exact evaluations using the physical models are computed. The comparison is performed considering: a quadratic approximation using LS method; neural networks. For this specific problem we have relevant advantages using neural networks. The mean absolute value of error and also the standard deviation of the absolute value of the error are shown. We can assume that neural networks constitute nowadays one of the more flexible method to approximate the complex relationships between design variables and performance indexes.

---