I´m running some simulation experiments and I need to determine the simulation run-length period. When I used others simulations software, they have the method batch means that do it for me.
I read the Plant Simulation documentation and I did not found how to do it in Plant Simulation. My Plant is the v. 12.2 with standard licence.
Anybody have some suggestion?
I mean simulation run-length is the time duration of the simulation.
You determine it in the EventController -> Settings -> Start/End.
The problem is how long I need to run the simulation to have statistically independence of my performance measures and calculate the confidence interval of my means, excluding the warm-up period.
I was in a similar discussion some time ago, although it concerned the warm-up period. You can find it here, together with an example: http://community.plm.automation.siemens.com/t5/Plant-Simulation-Forum/maximum-warm-up-period/m-p/337...
As you can see, the theory aroudn this topic tends to be a bit outdated, but it remains relevant for the critical cases. And, of course, you need to justify your choices regarding these settings.
The batch-means procedure is also the assumption of the ExperimentManager (although you are not necessarily restricted to it). Typically, you will first determine the warm-up period, then the run-length, then the number of observations. But the run-length often also relates to some practical figure, like a full month of production. You can play a bit with the latter two, to get reasonable CPU times.
There is also a tool inside Plant Simulation that helps you to determine the number of observations, it is called Sequential Sampler and you can load it through the button Manage Class Library, second tab (Libraries).
Hope this helps!
Thanks for your answer.
I will try what you suggest, but I´m afraid that because my Plant Simulation uses the standard licence, I can´t use the experiment manager.
But, it is what I need now. Define the warm-up, justify the run-lenght and define the number of observations. In my case, it is necessary explain or justify my choices and, if the Plant Simulation do it for me, its easier to do that.
Later, I will post here if it worked.
Since quite some versions the ExperimentManager is part of the Standard license. If it is not part of your model, you can load it just liek the Sequential Sampler.
If you have more questions, feel free to ask!
Hello Nomden and community,
I tried to use the Experiment Manager, but I couldn´t figure out how I could determine the run length period of the simulation with that. I only calculate the confidence intervals.
One strategy is to determine the number of observations, find the slower process to be monitored and then, calculate how long I need to run the simulation to capture the number of observation needed of this process.
To do that, I found in the documentation of the Plant Simulation (and suggestion of Nomden) the Sequential Sampler. But, unfortunatelly, it is not available to the standard licence.
I did my simulation experiments but I cannot justify the run-length that I choosed. It seems ok, but the report I will do needs a clear justification of my choices, because it will be read by people with scientific background.
Anybody has any suggestion?
The purpose of the a) warm-up period, b) run-length and c) sample size is to get rid of biases in your measurements. Obviously, you first need to determine a) and it seems you succeeded.
Then, to achieve the same statistical accuracy, you should realize that there is an interaction between b) and c). If you have shorter runs, you need more samples and vice versa. From here you have to make some practical considerations.
So, how to proceed? You should actually do some test runs, with the warm-up period in place!
Here you can still use the ExperimentManager to determine the run-length. You should set the number of replications (e.g. to 5 or 10) and then vary the run-length for a few of the worst-case experiments. By plotting the confidence intervals, you can see where the cumulative means stabilize and how it impacts the width of the conmfidence interval. Textbooks (Simulation: the practice of model development and use, by Stewart Robinson) do describe these considerations.
Most of the time, you can choose run-length and sample size on the conservative side, since the ExperimentManager does all the work. It can run overnight and -even better- during the weekend ("I worked all weekend to get the results out")!