morris

Screening with the Morris method

This market entry shows how to use the Morris sensitivity analysis method for the screening of a model in OpenMOLE.

Several OpenMOLE workflows are proposed here:

• "MorrisVerify.oms" is provided so anyone can ensure the method works as expected (see below for details)
• "MorrisTraffic.oms" applies the Morris sensitivity analysis to the netlogo Traffic model
• "MorrisTrafficWithPlot.oms" shows an example of using R to plot the standard Morris diagram

Verify

"MorrisVerify.oms" is provided so anyone can ensure the implementation of the method works as expected. It also stands as a good starting point to understand the intuition of the method.

In this simple workflow, there are 5 float inputs a,b,c,d and e.

Three functions are evaluated:

x = 2b + 3c - d^2 + e^3
y = 10a + 2bc + 3bcd
z = a

At execution time, results are stored into a subdirectory named "results_verify"

function x is designed such as variables should no lead to indirect effect, so impact of sigma(anything,x)=0 (except for powers)

A sensitivity analysis should detect as a result:

• a has no direct impact on x (mu=0) and no indirect effect through another variable (sigma=0)
• b has a direct impact on x (mu=2b) and no indirect effect (sigma=0)
• c has a direct impact on x (mu=3c) and no indirect effect (sigma=0)
• d has a strong impact on x (mu=d^2) and an indirect effect due to the power (sigma>0)
• e has a strong impact on x and an indirect effect due to the power (sigma>0)

function y should lead to strong interactions between some variables:

• a has a direct impact (mu=a) and no indirect effect (sigma 0)
• b has direct impact and an indirect effect (sigma >> 0)
• c has a direct impact which also depends on another variable (sigma >> 0)
• d has a direct impact and also depends on another variable (sigma >> 0)

A good way to understand the role of levels and repetitions count is to reproduce this experiment with different parameters.

Traffic Basic

The "MorrisTraffic.oms" workflow applies the Morris method to the Traffic Basic NetLogo model This model is unchanged, except to make it compliant with OpenMOLE.

However, as this model is stochastic, the results might be kind of unstable.

Note the impact of changing the parameters for levels and repetitions. Also notice how, when the parameter verbose = true is used, messages in the simulation results explain how the computation of the indicators is made.

Plot graph

Morris analysis usually is analyzed with graphs showing, for each input, its impact on all the outputs of the model on the two indicators mu and sigma. Outputs with an impact on mu have a direct impact; if sigma is huge, this impact is more dependant on other variables.

Creating the plots of a method is considered as a bad practice in OpenMOLE. It's better to download the CSV file and plot it by yourself, so you can adapt the graph to your needs, your precise experiment, etc. On another hand, screening is supposed to be quick to drive; reproducible research also recommands an automatic chain for building the graphs.

Should you desire a quick plot of the results, you might have a look to the "MorrisTrafficWithPlot.oms" workflow. In this workflow, we add an RTask which loads the outputs of the sensitivity analysis and relies on R to generate the 2D plots. The underlying R script is generic enough to adapt to your inputs and outputs. The R script might constitute a starting point for plotting your own graphs on your computer.