AUTOMATING WARM-UP LENGTH ESTIMATION Kathryn Hoad Stewart Robinson Ruth Davies Warwick Business School The University of Warwick Coventry, UK ABSTRACT There are two key issues in assuring the accuracy of esti- mates of performance obtained from a simulation model. The first is the removal of any initialisation bias, the sec- ond is ensuring that enough output data is produced to ob- tain an accurate estimate of performance. This paper is concerned with the first issue, and more specifically warm- up estimation. A continuing research project is described that aims to produce an automated procedure, for inclusion into commercial simulation software, for estimating the length of warm-up and hence removing initialisation bias from simulation output data. 1 INTRODUCTION Initialisation bias occurs when a model is started in an ‘un- realistic’ state. The output data collected during the warm- ing-up period of a simulation can be misleading and bias the estimated response measure. The removal of initialisa- tion bias is, therefore, important for obtaining accurate es- timates of model performance. Initialisation bias occurs primarily in non-terminating simulations, but in some instances it can also occur in ter- minating simulations. For instance, if a week’s production schedule is simulated it would be wrong to assume that there is no work-in-progress on the Monday morning. If we were to simulate the lunch time period of a shop it would be wrong to ignore the customers who may already be in the shop at the start of the period of interest. There are five main methods for dealing with initiali- sation bias (Robinson 2004): 1. Run-in model for a warm-up period until it reach- es a realistic condition (steady state for non- terminating simulations). Delete data collected from the warm-up period. 2. Set initial conditions in the model so that the si- mulation starts in a realistic condition. 3. Set partial initial conditions then warm-up the model and delete warm-up data. 4. Run model for a very long time making the bias effect negligible. 5. Estimate the steady state parameters from a short transient simulation run (Sheth-Voss et al. 2005). This project uses the first method; deletion of the data with initial bias by specifying a warm-up period (trunca- tion point). The key question is “how long a warm-up pe- riod is required?” The overall aim of the work is to create an automated procedure for determining an appropriate warm-up period that could be included in commercial si- mulation software. This paper describes the work that has been carried out to date with the aim of producing an automated procedure to estimate the warm-up period. Section 2 describes the extensive literature review that was carried out to find the various warm-up methods in existence. Section 3 explains how we short listed candidate methods for further testing. The next two sections describe the testing procedure, in- cluding the creation of artificial data sets and performance criteria. Sections 6 sets out the test results and Section 7 contains the summary and conclusions including plans for future work. 2 LITERATURE REVIEW An extensive literature review of warm-up methods was carried out in order to collect as many published methods and reviews of such methods as possible. 2.1 Warm-up methods in literature Through the literature search we found 42 warm-up meth- ods. Each method was categorised into one of 5 main types of procedure as described by Robinson (2004): 532 978-1-4244-2708-6/08/$25.00 ©2008 IEEE Proceedings of the 2008 Winter Simulation Conference S. J. Mason, R. R. Hill, L. Mönch, O. Rose, T. Jefferson, J. W. Fowler eds.