Even though we may be able to verify the statistical results of a simulation model, we still may not know whether the model actually replicates what is going on in the real world. The user of simulation generally wants to be certain that the model is internally correct and that all the operations performed in the simulation are logical and mathematically correct. An old adage often associated with computer simulation is "garbage in, garbage out." To gain some assurances about the validity of simulation results, there are several testing procedures that the user of a simulation model can apply.
Simulation models must be validated to make sure they are accurately replicating the system being simulated .
First, the simulation model can be run for short periods of time or for only a few simulation trials. This allows the user to compare the results with manually derived solutions (as we did in the examples in this chapter) to check for discrepancies. Another means of testing is to divide the model into parts and simulate each part separately. This reduces the complexity of seeking out errors in the model. Similarly, the mathematical relationships in the simulation model can be simplified so that they can more easily be tested to see if the model is operating correctly.
Sometimes manual simulation of several trials is a good way to validate a simulation .
To determine whether the model reliably represents the system being simulated, the simulation results can sometimes be compared with actual real-world data. Several statistical tests are available for performing this type of analysis. However, when a model is developed to simulate a new or unique system, there is no realistic way to ensure that the results are valid.
An additional problem in determining whether a simulation model is a valid representation of the system under analysis relates to starting conditions. Should the simulation be started with the system empty (e.g., should we start by simulating a queuing system with no customers in line), or should the simulation be started as close as possible to normal operating conditions? Another problem, as we have already seen, is the determination of how long the simulation should be run to reach true steady-state conditions, if indeed a steady state exists.
In general, a standard, foolproof procedure for validation is simply not possible. In many cases, the user of a simulation model must rely on the expertise and experience of whoever develops the model.