Simulation models are integral to modern scientific research, national defense, industry and manufacturing, and in public policy debates. These models tend to be extremely complex, often with thousands of factors and many sources of uncertainty. To understand the impact of these factors and their interactions on model outcomes requires efficient, high-dimensional design of experiments. Unfortunately, all to often, many large-scale simulation models continue to be explored in ad hoc ways. This suggests that more simulation researchers and practitioners need to be aware of the power of experimental design in order to get the most from their simulation studies. In this tutorial, we demonstrate the basic concepts important for design and conducting simulation experiments, and provide references to other resources for those wishing to learn more. This tutorial (an update of previous WSC tutorials) will prepare you to make your next simulation study a simulation experiment.
INTRODUCTIONIn June 2008, a new supercomputer called the "Roadrunner" was unveiled. This bank of machines was assembled from components originally designed for the video game industry; it costs $133 milion, and is capable of doing a petaflop (a quadrillion operations per second). The New York Times coverage stated that "petaflop machines like Roadrunner have the potential to fundamentally alter science and engineering" by allowing researchers to "ask questions and receive answers virtually interactively" and "perform experiments that would previously have been impractical" (Markoff 2008). Four years later, IBM's "Sequoia" supercomputer is the new world leader, with 16 petaflop capability. Yet let's take a closer look at the practicality of a brute-force approach to simulation experiments. Suppose a simulation has 100 factors, each factor has two levels (say, low and high) of interest, and we decide to look at each combination of these 100 factors. A single replication of this experiment for simulation that runs as fast as a single operation would take over 2.5 million years on the Sequoia and over 40 million years on the Roadrunner! Efficient design of experiments can break this curse of dimensionality at a tiny fraction of the cost. For example, suppose we want study 100 factors and all their two-way interactions. One screening design we could use (a resolution 5 fractional factorial, described in Section 3.3) specifies 32768 specific combinations of the factor levels to evaluate. How quickly can we finish such an experiment? On a desktop computer with a simulation that takes a full second to run, each replication of this experiment takes under 9.5 hours; even if the simulation takes a more reasonable one minute to run, we can finish this experiment on an 8-core desktop (under $3,000) in 2.85 days. Other designs are even more efficient, and provide more detailed insights into the simulation model's behavior.The field called Design of Experiments (DOE) has been around for a long time. Many of the classic experimental designs can be used in ...