Steady state simulation

[farnaz]

Hi,

In my simulation model, I am trying to reach the steady state for the results. So, I run the model with 100 replications for 5 year but if I change the 5 years to 6 years, or consider a 1 year warm up time, or 2 years warm up time for the system the results will vary differently for each changes . The changes are small for example the total waiting time increase around 4 minutes when I consider 2 years warm up for the system. So, how can I make sure the simulation results are steady state?

[dsturrock]

The primary objective of setting warmup time is to reduce or eliminate the effects of any startup bias from the reported statistics (especially when starting "idle & empty"). Perhaps the best way to determine an appropriate warmup time is by plotting the observations of some key statistics and then visually determining at what point the starting bias is minimized. You can accomplish this by:

--Creating a real-time plot in the facility view.

--Enable logging on some Tallies and then create a Dashboard on that log.

--Use a Write step or ExcelWrite step to write each Tally observation to a file and plot it externally.

Generally an approximation is "good enough".

--If you make it too short, you may include some bias - but the further away you are from time 0 the less the impact of that bias.

--If you make it too long, you are just throwing away potentially good results requiring a longer replication to get equivalent results.

A warmup that is too long is safer than a warmup that is too short.

In general, the longer your replication is compared to the warmup time, the less important the warmup is. For example if you are running for 8 hours and it requires about an hour to reach steady state, then the warmup period is quite important since more than 10% of your data is potentially biased. If you are running for a year, and you reach steady state in a few hours, then the warmup period is much less important and possibly even unnecessary since less than 0.1% of your data is potentially biased.

The number of replications you need to run is a totally different issue and has no impact on the warmup time issue.

Any good simulation textbook, such as Simio and Simulation: Modeling, Analysis, Applications should explain this in detail.