ON VARIABLE SELECTION AND EFFECTIVE ESTIMATIONS OF INTERACTIVE AND QUADRATIC SENSITIVITY COEFFICIENTS: A COLLECTION OF REGULARIZED REGRESSION TECHNIQUES

摘要

In this paper, we present effective regularized regressions on variable selection and estimation of sensitivity coefficients of k_(eff) of a TRIGA fuel pin model, up to Order 2. 23 parameters are regarded. Considering 253 interactive and 23 quadratic terms, there are 299 parameters in total. Parameters were sampled via Latin Hypercube sampling design. We explored the variable selection among the 299 parameters with seven types of regularized methods with different sample sizes and found several methods, e.g. Bayesian lasso, Bayesian ridge, etc. outperform the commonly used lasso compared with the reference. Also, we compared these methods, including lasso, with linear regression on sensitivity estimation with 299 realizations. Result showed the effectiveness of several regularized methods, e.g. Bayesian ridge, on estimating high order coefficients compared with the reference from forward simulations. When reducing sample size, results show that some methods can still estimate interactions acceptably well.