On Improved Least Squares Regression and Artificial Neural Network Meta-Models for Simulation via Control Variates
Abstract
The analysis of large-scale simulations can pose a large computational burden, often necessitating the use of high performance computers. Moreover, these models can be analytically intractable because of complex, internal logical relationships. To begin to overcome these types of obstacles, a method known as meta-modeling has been developed to construct mathematical functions that act as analytic surrogates to large scale simulations. This research examines the introduction of second-order interactions for two types of asymptotically-standardized linear control variates to least squares regression and radial basis neural network meta-models for a queuing simulation. Extensions are made to the statistical framework for variance reduction of direct estimation of single response, single population simulations and the framework for meta-models of single response, multiple population simulations. As a result, the new extensions are shown to significantly outperform existing frameworks and also provide the means to interpret and better understand the system dynamics that manifest when a system exhibits an exceptional amount of variance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 15, 2016
- Accession Number
- AD1017989
Entities
People
- Michael P. Gibb
Organizations
- Air Force Institute of Technology