Machine learning for parameter screening in computer simulations

Abstract

The aim of this paper is to highlight the potential of Machine Learning for parameter screening in computer simulations, presenting alternative approaches to automatic parameter ranking and screening. This is indeed a fundamental step in the development of a simulator, because it allows reducing the dimensionality of the parameter set, making model tuning more efficient. With parameter ranking we denote the process of measuring the relevance of the parameters for accurately simulating a phenomenon, while with parameter screening we denote the choice of a specific subset of parameters to be used for model tuning. We will present ranking techniques based on Logistic Regression and Multilayer Perceptron, and a simple procedure for going from ranking to screening. Our techniques have been validated against a helicopter simulator case-study but the techniques do not rely on any domain-specific feature or assumption.