Machine Learning Approaches for Computing Information Value and Information Density

Information value (VoI) analysis is a key component for decision-making supported by quantitative simulations. A key obstacle to the full utilization of VoI is computational efficiency. This thesis examines the use of machine learning approaches to estimating VoI for realistic simulators. We compare the smoothing approaches already introduced, and propose two novelties. First, an approach based on the nearest neighbors. We prove a central limit theorem and then we discuss the automatic selection of the number of neighbors through a LassoLars weighting approach. We also propose a modification of a previously introduced algorithm by using MARS regression. We compare the resulting estimators through a wide range of numerical experiments. We then adapt the algorithms for the estimation of a new quantity, the information density. Experiments show that the algorithms can be successfully modified and one obtains consistent indications about the regional importance of variables, both individually and in groups.