© J.C. Baltzer AG, Science Publishers
On performance potentials and conditional Monte Carlo
for gradient estimation for Markov chains
Xi-Ren Cao
a
, Michael C. Fu
b
and Jian-Qiang Hu
c
a
Department of Electrical Engineering,
The Hong Kong University of Science and Technology,
Clear Water Bay, Kowloon, Hong Kong
b
The Robert H. Smith School of Business, Institute for Systems Research,
University of Maryland, College Park, MD 20742, USA
c
Department of Manufacturing Engineering, Boston University,
Boston, MA 02215, USA
We consider the problem of sample path-based gradient estimation for long-run (steady-
state) performance measures defined on discrete-time Markov chains. We show how two
estimators – one derived using the likelihood ratio method with conditional Monte Carlo
and splitting, and the other derived using performance potentials and perturbation analysis –
are related. In particular, one can be expressed as the conditional expectation of a suitably
weighted average of the other. This demonstrates yet another connection between the two
gradient estimation techniques of perturbation analysis and the likelihood ratio method.
Keywords: Markov chains, perturbation analysis, likelihood ratio method, gradient esti-
mation, potential theory
1. Introduction
Many stochastic systems can be modeled by Markov chains with a finite state
space. However, due to the complexity of realistic systems of interest, the resulting
model may have an enormous state space that leads to intractability in terms of com-
putational complexity. Large closed queueing networks are a classic example of this
phenomenon. In this case, it is often more computationally efficient to simulate the
system rather than to try to solve it analytically. In order to perform sensitivity analysis
or optimization when simulation is employed, various gradient estimation techniques
have been developed over the past two decades, most notably the techniques of pertur-
bation analysis (cf. Ho and Cao [14], Glasserman [13], Cao [1] and Fu and Hu [11])
and the likelihood ratio method, which is also sometimes referred to as the score
function method (cf. Rubinstein and Shapiro [15]); see also Fu [10] for a review of
Annals of Operations Research 87(1999)263–272 263