Model selection using an information theory approach

In this thesis we use the information theoretic approach in selecting the best model among many candidate models. It is shown that the information theoretic approach is better than the standard R2 approach in selecting models. We use Akaike Information Criteria (AIC) to select the best model for resilient modulus of a soil and for a girder. This approach is applied to statistical models, neural network models and physics based models. The information theory approach is compared with the R2 approach and it is found that the information theoretic approach is more stable and gives better results. The notion of ranking stability is introduced and is used as one of the reasons that makes information theory approach better than the R2 approach. Important results are captured and compared to the results of the R 2 method in two different data sets.