**6. Conclusions**

This work has described an approach for obtainment and formal validation of LRMs, by means of the combination of genetic programming with statistical models. Our approach used the Audze-Eglais Uniform Latin Hypercube technique for the selection of samples with high representative power to form the training set. In order to evaluate the LRMs found with the introduced technique, we used statistical tests of hypothesis and residual analysis, aiming to verify the assumptions about the structures of the errors of these models.

In order to validate the proposed approach, we used a case study, with the prediction of performance in embedded systems. The problem of the case study consisted in exploring the configurations of a data bus in order to optimize the performance of the embedded application of sorting a set of integers by radix. So, with the use of the proposed technique, we generated LRMs capable of estimating the performance for all of the bus configurations.

The validation stages allowed us to realize that the LRMs found are adequate to the prediction of performance of the application, since all the assumptions about the structures of the errors were verified. So, the final LRMs were able to estimate the performances accurately, presenting mean global errors below 5%.
