INTRODUCTION {sensitivity}R Documentation

Package 'sensitivity': Sensitivity Analysis

Description

The sensitivity package implements sensitivity analysis methods: linear and monotonic sensitivity analysis (SRC, PCC, SRRC, PRCC), the screening method of Morris, and non-linear global sensitivity analysis (the Sobol indices, the FAST method).

The functions of this package generate the design of experiments (depending on the method of analysis) and compute the sensitivity indices based on the model inputs and outputs. All sensitivity indices can be estimated with the bootstrap technique which allows to estimate the bias, and basic bootstrap confidence intervals. Text and graphical outputs display the results of the analysis.

Details

The approach applied when performing a sensitivity analysis (SA) is as follows:

step 1
The model is defined: it is a function that returns the (real) ouput values (called responses), corresponding to a sample of (real) input parameters (called factors).
step 2
A sensitivity analysis method is chosen. Parameters of this method must be in accordance with objectives and technical constraints (like computational time).
step 3
A design of experiments (DOE) corresponding to the SA method is generated.
step 4
The model is evaluated on the DOE values.
step 5
The sensitivity indices are computed, based on input and output values.
step 6
Post-treatments...

The sensitivity package allows to follow this methodology:

(step 1) The model can be internal or external to R. If internal, it can be a function that takes an unique matrix or data.frame parameter and returns a numeric vector. It can also be a predictor, i.e. an object wich can be called with the predict method. One should note that all the responses must be computed by a single call to the model function (then, the model can be vectorized).
If the model is external it does not have to be interfaced with R: the user won't have to give a model to the function. Then, it will stop just after generating the DOE. The responses have to be computed by the user, whithin R or not. Calculations will start again when the user gives the corresponding responses (via the tell function).

The four next steps depend upon the type of the model:

For internal models:
(step 2-5) sa <- method(model, parameters...)

For external models:
(step 2-3) sa <- method(model = NULL, parameters...)
(step 4) external to R (or not), and the result is loaded by the user in the y variable
(step 5) tell(sa, y)

method should be the name of a SA function, such as srcpcc, morris, sobol, or fast. These function create the object sa of class "srcpcc", "morris", "sobol", or "fast". For further information on these function, see the corresponding documentation.

Finally, for displaying the results of the analysis:
(step 6) print(sa); plot(sa)

References

Saltelli, A., Chan, K. and Scott, E. M., 2000, Sensitivity analysis. Wiley, 33–45.

See Also

srcpcc morris sobol fast tell testmodels


[Package sensitivity version 1.3-0 Index]