On Initial Transient Detection in the Setting of Steady-state Simulation

Abstract

In the setting of performance engineering applications of simulation, the most widely used performance measures relate to the steady-state of the system being simulated. However, computing a steady-state expectation is often challenging, both because of the presence of the initial transient due to initialization via a distribution which is generally atypical of equilibrium behavior, and because of the difficulties inherent in building confidence intervals in the presence of auto-correlated data. This was a problem that was first addressed via the ratio estimator introduced by Iglehart et al. in the 1970 and discussed in detail in Reuven Rubinstein 1981 book Simulation and the Monte Carlo Method. In this talk, we will discuss several new sampling-based algorithms for mitigating the presence of the initial transient that have a provable theoretical basis, and will also discuss some theoretical limitations of existing algorithms for constructing steady-state confidence intervals that are likely to have an impact on their practical performance. In contrast to much of the existing theoretical literature on the initial transient problem, our methods rely on using statistical information accumulated from the simulation itself to correct for the presence of the initial transient, rather than using theoretical bounds on rates of convergence or carefully crafted perfect simulation algorithms. This talk is based on joint work with Hernan Awad and Jose Blanchet.