Resumo:
Computer model validation of Discrete Event Simulation (DES) is essential for project success since this stage guarantees that the simulation model corresponds to the real system. Nevertheless, it is not possible to assure that the model represents 100% of the real system. The literature suggests using more than one validation technique, but statistical tests are preferable. However, they have limitations, since the tests usually test the mean or standard deviation individually, and do not consider that the data may be within a pre-established tolerance limit. In this way, Generative Adversarial Networks (GANs) can be used to train, evaluate and discriminate data and validate DES models, because they are two competing neural networks, where one generates data and the other discriminates them. The proposed method is divided into two phases. The first is the "Training Phase" and it aims to train the data. The second, the "Test Phase" aims to discriminate the data. In addition, in the second phase, the Equivalence Test is performed, which statistically analyze if the difference between the judgments is within the tolerance range determined by the modeler. To validate the proposed method and to verify the Power Test, experiments were carried out in continuous, discrete, and conditional distributions and in a DES model. From the tests, the Power Test curves were generated considering a real tolerance of 5.0%, 10.0% and 20.0%. The results showed that it is more efficient to use the dataset that presents larger sample in the “Test Phase” while the set with smaller sample size needs to be used in the “Training Phase”. In addition, the confidence of the Power Test increases with big higher dataset in first phase, presenting smaller confidence intervals. Also, the more metrics are evaluated at once, the greater the amount of data inputted in the GANs' training. The method suggests classifying a validation based on the achieve tolerance: Very Strong, Strong, Satisfying, Marginal, Deficient and Unsatisfying. Finally, the method was applied to three real models, two of them in manufacturing and the last one in the health sector. We conclude that the proposed method was efficient and was able to show the degree of validation of the models that represent the real system.