sensitivity analysis (vensim)

[wanderbeg]

Hi, guys

I have this difficulty.
I am using the vensim DSS, and I plan to do sensitivity analysis. I can do this in the Monte Carlo function, but when I try generate the graph, I can not, that is, the sensitivity analysis button in the vensim does not appear.
What can it be?

Best regards,

[tomfid]

Try loading a new Analysis Toolset?

[wanderbeg]

Try loading a new Analysis Toolset?

Where can I do this?

[Administrator]

Click Tools->Analysis Toolset->Modify.

[tomfid]

Or just Tools>Analysis>New to reload a new copy of the toolset.

[wanderbeg]

Thanks, guys!

However, I have another problem. The graphs of the sensitivity analysis were different and did not appear those indicators of sensitivity 50%, 57%, 95% and 100%.
I have annexed!

What do you think?

[tomfid]

Right-click the sensitivity graph tool and change the options to show confidence bounds rather than individual traces.

From your second graph, it looks like the range of behavior is very narrow.

[wanderbeg]

I simulate my parameters, but in all graphs the intervals are constant.
Estranged!

[tomfid]

What is shown on this plot? If it's a parameter, it will be constant.

[tomfid]

Try the samples in the help system: http://vensim.com/documentation/usr15.htm

[wanderbeg]

Thank you for your help!

[RWTH_FCN]

Hi,

I have two questions regarding sensitivity analysis and stochastic programming in Vensim:

1- Can we do these sentivity analysis in a model which has external function? I am asking this becasue I get this errors:

{ERROR: Yearly Depreciation Factor not found or wrong type must be constant.
ERROR: No sensitivity inputs marked in sensitivity file}


2- I have a large-scale model for electricity market, which minimizes the total system cost (I have an external function for this minimization function). I want to include uncertainty in the model and consider some parameters (assume fuel price) uncertain in order to model the system stochastically. I was wondering how could I solve my problem stochastically (if its possible) with facilities of Vensim like Payoff definition or Optimization options when I have some external functions as well?

Best Regards

[Administrator]

1- Can we do these sentivity analysis in a model which has external function? I am asking this becasue I get this errors:

{ERROR: Yearly Depreciation Factor not found or wrong type must be constant.
ERROR: No sensitivity inputs marked in sensitivity file}

These errors are nothing to do with external functions. I'd suggest looking in the help system first and trying the example there before trying it on your own model.

[tomfid]

Sounds like your control file (.vsc) specifies a variable that doesn't exist or isn't constant.

[RWTH_FCN]

1- Can we do these sentivity analysis in a model which has external function? I am asking this becasue I get this errors:

{ERROR: Yearly Depreciation Factor not found or wrong type must be constant.
ERROR: No sensitivity inputs marked in sensitivity file}

These errors are nothing to do with external functions. I'd suggest looking in the help system first and trying the example there before trying it on your own model.

I did it with a sample model W/O external function and it was okay, that's why I was wondering can external functions cause any failure for sensitivity analysis?

[RWTH_FCN]

Sounds like your control file (.vsc) specifies a variable that doesn't exist or isn't constant.

The variable is constant.
I would be pleased to have your feedback about my second question as well.

[tomfid]

I think the second error is a consequence of the first. It's not finding Yearly Depreciation Factor, and therefore not adding it as an input. Are there other parameters in the vsc?

Does Yearly Depreciation Factor have an array dimension? It's not set to "Initial" type by accident?

[tomfid]

I think the "wrong type" error only occurs if a variable is not a constant (the "not found" is misleading). Most likely causes are setting the constant subtype to unchangeable (with ==), setting the type to initial(), or simply specifying a nonnumeric expression for the variable, like x*y or 3/4.

[RWTH_FCN]

I found it. The reason was that the variable had subscripts. So I used another variable W/O subscripts and it worked well.
Can I ask you how can I do sensivity analysis for variables with subscripts and variables with array dimension?

[RWTH_FCN]

I found it. The reason was that the variable had subscripts. So I used another variable W/O subscripts and it worked well.
Can I ask you how can I do sensivity analysis for variables with subscripts and variables with array dimension?

I solved it too. I should have just included subscripts in the .vsc file.

[RWTH_FCN]

I have a large-scale model that its simulation time is 7 hours (I am using compiler as well). From computational points of view, is it possible to do a Monte carlo simulation on it? Is there any linear relation between "number of simulation" in Monte carlo tab and "time of simulation" in vensim?

[Administrator]

You can run monte-carlo, but it will probably take you years to run it if a single run takes 7 hours. Each monte-carlo simulation will take 7 hours.

[RWTH_FCN]

You can run monte-carlo, but it will probably take you years to run it if a single run takes 7 hours. Each monte-carlo simulation will take 7 hours.

Thanks. Do you have any suggestion for modeling uncertainty in such models which take very long in Vensim? Is there any sample model for simulating uncertainty with feedback-loops in Vensim itself?

[tomfid]

There's a little advice on speed here: https://metasd.com/2011/01/optimizing-vensim-models/
One item that's not mentioned there is that restricting the savelist helps.

We have an experimental version that does parallel multicore execution, which would reduce your simulation time by a factor of the number of cores you have (I have a server with 60, so it does in a minute what normally takes an hour). We could explore that option, though memory might be a problem. Still, doing 1000 or 1000000 runs would be a little daunting.

[RWTH_FCN]

We have an experimental version that does parallel multicore execution, which would reduce your simulation time by a factor of the number of cores you have (I have a server with 60, so it does in a minute what normally takes an hour). We could explore that option, though memory might be a problem. Still, doing 1000 or 1000000 runs would be a little daunting.

Great. Have you released that version or is it possible to have it in the near future? Using such version can speed up simulation of large-scale model noticeably.