ODEsobol.ODEnetwork {ODEsensitivity} | R Documentation |
Sobol' Sensitivity Analysis for Objects of Class ODEnetwork
Description
ODEsobol.ODEnetwork
performs the variance-based Sobol' sensitivity
analysis for objects of class ODEnetwork
. Package
ODEnetwork
is required for this function to work.
Usage
## S3 method for class 'ODEnetwork'
ODEsobol(mod, pars, times, n = 1000, rfuncs = "runif",
rargs = "min = 0, max = 1", sobol_method = "Martinez",
ode_method = "lsoda", parallel_eval = FALSE, parallel_eval_ncores = NA,
...)
Arguments
mod |
[ |
pars |
[ |
times |
[ |
n |
[ |
rfuncs |
[ |
rargs |
[ |
sobol_method |
[ |
ode_method |
[ |
parallel_eval |
[ |
parallel_eval_ncores |
[ |
... |
further arguments passed to or from other methods. |
Details
If the object of class ODEnetwork
supplied for mod
doesn't
include any events, the solution of the ODE network is determined
analytically using simuNetwork
. In the presence
of events, simuNetwork
uses
ode
to solve the ODE network numerically.
The sensitivity analysis is done for all state variables and all
timepoints simultaneously. If sobol_method = "Jansen"
,
soboljansen
from the package sensitivity
is used to estimate the Sobol' sensitivity indices and if
sobol_method = "Martinez"
, sobolmartinez
is used (also from the package sensitivity
).
Value
List of length 2 * nrow(mod$state)
and of class
ODEsobol
containing in each element a list of the Sobol' sensitivity
analysis results for the corresponding state variable (i.e. first order
sensitivity indices S
and total sensitivity indices T
) for
every point of time in the times
vector. This list has an extra
attribute "sobol_method"
where the value of argument
sobol_method
is stored (either "Jansen"
or
"Martinez"
).
Note
In situations where the solution of the ODE model has to be determined
numerically, it might be helpful to try a different type of ODE-solver
(argument ode_method
) if the simulation of the model takes too long.
The ode_method
s "vode"
, "bdf"
, "bdf_d"
,
"adams"
, "impAdams"
and "impAdams_d"
might be faster than the standard ode_method
"lsoda"
.
If n
is too low, the Monte Carlo estimation of the sensitivity
indices might be very bad and even produce first order indices < 0 or
total indices > 1. First order indices in the interval [-0.05, 0) and total
indices in (1, 1.05] are considered as minor deviations and set to 0
resp. 1 without a warning. First order indices < -0.05 or total indices
> 1.05 are considered as major deviations. They remain unchanged and a
warning is thrown. Up to now, first order indices > 1 or total indices < 0
haven't occured yet. If this should be the case, please contact the package
author.
Author(s)
Frank Weber
See Also
soboljansen,
sobolmartinez,
plot.ODEsobol
Examples
##### A network of 4 mechanical oscillators connected in a circle #####
# Definition of the network using the package "ODEnetwork":
M_mat <- rep(2, 4)
K_mat <- diag(rep(2 * (2*pi*0.17)^2, 4))
K_mat[1, 2] <- K_mat[2, 3] <-
K_mat[3, 4] <- K_mat[1, 4] <- 2 * (2*pi*0.17)^2 / 10
D_mat <- diag(rep(0.05, 4))
library("ODEnetwork")
lfonet <- ODEnetwork(masses = M_mat, dampers = D_mat, springs = K_mat)
# The parameters to be included in the sensitivity analysis and their lower
# and upper boundaries:
LFOpars <- c("k.1", "k.2", "k.3", "k.4",
"d.1", "d.2", "d.3", "d.4")
LFObinf <- c(rep(0.2, 4), rep(0.01, 4))
LFObsup <- c(rep(20, 4), rep(0.1, 4))
# Setting of the initial values of the state variables:
lfonet <- setState(lfonet, state1 = rep(2, 4), state2 = rep(0, 4))
# The timepoints of interest:
LFOtimes <- seq(25, 150, by = 2.5)
# Sobol' sensitivity analysis (here only with n = 500, but n = 1000 is
# recommended):
set.seed(1739)
# Warning: The following code might take very long! There are warnings
# occurring which might be due to "n" being too low.
suppressWarnings(
LFOres_sobol <- ODEsobol(mod = lfonet,
pars = LFOpars,
times = LFOtimes,
n = 500,
rfuncs = "runif",
rargs = paste0("min = ", LFObinf,
", max = ", LFObsup),
sobol_method = "Martinez",
parallel_eval = TRUE,
parallel_eval_ncores = 2)
)