Initial Transient Period Detection Using Parallel Replications

This paper describes an extension of the replication/deletion approach for removal of the initialisation bias in the output data of discrete-event simulations. The proposed extension adaptively determines the cutoff index and also respects a user given ratio between the lengths of transient and steadystate period. We show the benefits of the extension inspecting the well-known M/M/1 queue. Using an example from the logistic area, we also point out that the proposed replication/deletion approach is additionally capable of detecting non-ergodic systems, which is a very difficult task for the usual single long-run simulation. The presented examples show that the extended replication/deletion approach leads to about the same qualitative steady-state estimates than single long-run simulations, but needs less time in practice.