mif
causes the iterated filtering algorithm to run for a specified number of iterations.
At each iteration, the particle filter is performed on a perturbed version of the model.
Specifically, parameters to be estimated are subjected to random perturbations at each observation.
This extra variability effectively smooths the likelihood surface and combats particle depletion by introducing diversity into the population of particles.
At the iterations progress, the magnitude of the perturbations is diminished according to a user-specified cooling schedule.
For most purposes, mif
has been superseded by mif2
.
"mif"(object, Nmif = 1, start, ivps = character(0), rw.sd, Np, ic.lag, var.factor = 1, cooling.type, cooling.fraction.50, method = c("mif","unweighted","fp","mif2"), tol = 1e-17, max.fail = Inf, verbose = getOption("verbose"), transform = FALSE, ...)
"mif"(object, Nmif = 1, Np, tol, ...)
"mif"(object, Nmif, start, ivps, rw.sd, Np, ic.lag, var.factor, cooling.type, cooling.fraction.50, method, tol, transform, ...)
"continue"(object, Nmif = 1, ...)
"conv.rec"(object, pars, transform = FALSE, ...)
"conv.rec"(object, ...)
pomp
.
ivps
must have a positive random-walk standard deviation specified in rw.sd
.
If there are no regular parameters with positive rw.sd
, i.e., only IVPs are to be estimated, see below "Using mif
to estimate initial-value parameters only".
names(rw.sd)
must be a subset of names(start)
,
The random walk is not dynamically added to the initial-value parameters (named in ivps
).
The algorithm requires that the random walk be nontrivial, so that rw.sd
be positive for at least one element.
Np
either as a vector of positive integers (of length length(time(object,t0=TRUE))
) or as a function taking a positive integer argument.
In the latter case, Np(k)
must be a single positive integer, representing the number of particles to be used at the k
-th timestep:
Np(0)
is the number of particles to use going from timezero(object)
to time(object)[1]
,
Np(1)
, from timezero(object)
to time(object)[1]
,
and so on, while when T=length(time(object,t0=TRUE))
,
Np(T)
is the number of particles to sample at the end of the time-series.
mif
update for initial-value parameters consists of replacing them by their filtering mean at time times[ic.lag]
, where times=time(object)
.
It makes no sense to set ic.lag>length(times)
;
if it is so set, ic.lag
is set to length(times)
with a warning.
rw.sd
.
In particular, the width of the distribution of particles at the start of the first mif
iteration will be random.walk.sd*var.factor
.
By default, var.factor=1
.
cooling.type
specifies the nature of the cooling schedule. When cooling.type="geometric"
, on the n-th mif
iteration, the relative perturbation intensity is cooling.fraction.50^(n/50)
.
When cooling.type="hyperbolic"
, on the n-th mif
iteration, the relative perturbation intensity is (s+1)/(s+n)
, where (s+1)/(s+50)=cooling.fraction.50
.
cooling.fraction.50
is the relative magnitude of the parameter perturbations after 50 mif
iterations.
method
sets the update rule used in the algorithm.
method="mif"
uses the iterated filtering update rule (Ionides 2006, 2011);
method="unweighted"
updates the parameter to the unweighted average of the filtering means of the parameters at each time;
method="fp"
updates the parameter to the filtering mean at the end of the time series.
pfilter
.
TRUE
, optimization is performed on the transformed scale, as defined by the user-supplied parameter transformations (see pomp
).
mif
returns an object of class mif
.
The latter inherits from the pfilterd.pomp
and pomp
classes.
mif
, while regular parameters are perturbed at the initial time and after every observation, IVPs are perturbed only at the initial time.mif
iterations, one can use the mif
method on a mif
object.
By default, the same parameters used for the original mif
run are re-used (except for tol
, max.fail
, and verbose
, the defaults of which are shown above).
If one does specify additional arguments, these will override the defaults.mif
iterations from where one left off using the continue
method.
A call to mif
to perform Nmif=m
iterations followed by a call to continue
to perform Nmif=n
iterations will produce precisely the same effect as a single call to mif
to perform Nmif=m+n
iterations.
By default, all the algorithmic parameters are the same as used in the original call to mif
.
Additional arguments will override the defaults.mif
's fixed-lag smoothing to estimate only initial value parameters (IVPs).
In this case, the IVPs to be estimated are named in ivps
and no positive entries in rw.sd
correspond to any parameters not named in ivps
.
If theta
is the current parameter vector, then at each mif
iteration, Np
particles are drawn from a normal distribution centered at theta
and with width proportional to var.factor*rw.sd
, a particle filtering operation is performed, and theta
is replaced by the filtering mean at time(object)[ic.lag]
.
Note the implication that, when mif
is used in this way on a time series any longer than ic.lag
, unnecessary work is done.
If the time series in object
is longer than ic.lag
, consider replacing object
with window(object,end=ic.lag)
.mif
object:
E. L. Ionides, A. Bhadra, Y. Atchad\'e, & A. A. King, Iterated filtering, Annals of Statistics, 39:1776--1802, 2011.
E. L. Ionides, D. Nguyen, Y. Atchad\'e, S. Stoev, and A. A. King. Inference for dynamic and latent variable models via iterated, perturbed Bayes maps. Proc. Natl. Acad. Sci. U.S.A., 112:719--724, 2015.
A. A. King, E. L. Ionides, M. Pascual, and M. J. Bouma, Inapparent infections and cholera dynamics, Nature, 454:877--880, 2008.
pomp
, pfilter
, mif2