Optimization Of Hyperparameters Of Anns – Application To The Second Virial Coefficient (b) Of Fluid Mixtures

Résumé: Machine learning is without a doubt one of the most powerful technologies in today's world, and its use has spread to reach all application and all fields. In machine learning building a good model is crucial to find a solution to a given problem and since hyperparameters have an important influence on the performance of the machine learning model therefore we must choose wisely its hyperparameters configuration. This thesis defines the hyperparameter optimization problem and its challenges and studies the use of several hyperparameter optimization methods on the problem of the prediction of the B coefficient of organic and inorganic mixtures, reflecting the PVT (pressure, volume, temperature) behaviour of fluids, using the fluids' critical properties, acentric factors and dipole moments as inputs. Moreover, this thesis offer various experiments comparing the performances of ten hyperparameter optimization methods such as Grid search, Bayesian Optimization, Optuna, etc. In addition we implemented a GUI that eases the usage of these methods, which makes it very easy for the non-expert to exploit these methods with only a basic knowledge of machine learning. The results obtained in this work show that using hyperparameters optimization methods have a high value and quality to find good configuration for our model.

Publié dans la revue: