Green Simulation: A Methodology for Reusing the Output of Past Computer Simulation Experiments

Northwestern University, Evanston IL

Investigators

Abstract

The standard practice in computer simulation experiments is to run an experiment to answer a question, and use the experiment's output only to answer that question. When future questions arise that are not answered adequately by output from past experiments, that output is not used at all in answering them. Instead, a new experiment is run, as though it were the first experiment run with that simulation model. This award will help make methods to make computer simulation experiments more efficient by reusing the output of old experiments. Simulation modeling is an important tool in military, business, science, and engineering applications. Computer simulation experiments often occupy scarce, expensive high-performance computing facilities and consume substantial amounts of electricity. If successful, this research will benefit society by reducing the resources consumed by computer simulation experiments. The project includes opportunities to train Ph. D. students and to integrate research findings into Ph. D. courses on simulation. Efforts will be made to recruit students from underrepresented groups; several female students have been part of the principal investigator's research group. The objective of this research is to improve the computational efficiency of stochastic simulation experiments in a setting in which there is a sequence of repeated experiments using the same simulation model with different inputs. Outputs of this research will include algorithms that have the potential to improve the efficiency of later experiments by storing and reusing the output of earlier experiments. The methods to be employed include the likelihood ratio method, metamodeling, and variance reduction techniques for stochastic simulation. This research is also applicable more broadly to statistics and analytics, not just to simulation. It will develop experimental designs that take into account the availability of data from previous experiments, determining what additional data may be acquired to answer the question at hand.