On-line error bounds for steady-state approximations: A potential solution to the initialization bias problem

By studying performance measures via reward structures, on-line error bounds are obtained by successive approximation. These bounds indicate when to terminate computation with guaranteed accuracy; hence, they provide insight into steady-state convergence. The method therefore presents a viable alternative to steady-state computer simulation where the output series is typically contaminated with initialization bias whose impact on the output cannot be easily quantified. The method is illustrated on capacitated queueing networks. The results indicate that the method offers a practical tool for numerically approximating performance measures of queueing networks. Results on steady-state convergence further quantify the error involved in analyzing an inherently transient system using a steady-state model