Data Farming the Parameters of Simulation-Optimization Solvers

The performance of a simulation-optimization algorithm, a.k.a. a solver, depends on its parameter settings. Much of the research to date has focused on how a solvers parameters affect its convergence and other asymptotic behavior. While these results are important for providing a theoretical understanding of a solver, they can be of limited utility to a user who must set up and run the solver on a particular problem. When running a solver in practice, good finite-time performance is paramount. In this paper, we explore the relationship between a solvers parameter settings and its finite-time performance by adopting a data farming approach. The approach involves conducting and analyzing the outputs of a designed experiment wherein the factors are the solvers parameters and the responses are assorted performance metrics measuring the solvers speed and solution quality over time. We demonstrate this approach with a study of the ASTRO-DF solver when solving a stochastic activity network problem and an inventory control problem. Through these examples, we show that how some of the solvers parameters are set greatly affects its ability to achieve rapid, reliable progress and gain insights into the solvers inner workings. We discuss the implications of using this framework for tuning solver parameters, as well as for addressing related questions of interest to solver specialists and generalists.

Data Farming the Parameters of Simulation-Optimization Solvers Academic Article