Commit ae72c409 authored by fred's avatar fred

first try for particular filter

git-svn-id: https://www.dynare.org/svn/dynare/trunk@3222 ac1d8469-bf42-47a9-8791-bf33cf982152
parent fe723f0d
function [LIK,lik] = monte_carlo_SIS_particle_filter(reduced_form_model,Y,start,number_of_particles)
% hparam,y,nbchocetat,nbchocmesure,smol_prec,nb_part,g,m,choix
% Evaluates the likelihood of a nonlinear model with a particle filter without systematic resampling.
%
% INPUTS
% reduced_form_model [structure] Matlab's structure describing the reduced form model.
% reduced_form_model.measurement.H [double] (pp x pp) variance matrix of measurement errors.
% reduced_form_model.state.Q [double] (qq x qq) variance matrix of state errors.
% reduced_form_model.state.dr [structure] output of resol.m.
% Y [double] pp*smpl matrix of (detrended) data, where pp is the maximum number of observed variables.
% start [integer] scalar, likelihood evaluation starts at 'start'.
% mf [integer] pp*1 vector of indices.
% number_of_particles [integer] scalar.
%
% OUTPUTS
% LIK [double] scalar, likelihood
% lik [double] vector, density of observations in each period.
%
% REFERENCES
%
% NOTES
% The vector "lik" is used to evaluate the jacobian of the likelihood.
% Copyright (C) 2009 Dynare Team
%
% This file is part of Dynare.
%
% Dynare is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% Dynare is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with Dynare. If not, see <http://www.gnu.org/licenses/>.
global M_ bayestopt_
persistent init_flag
persistent restrict_variables_idx observed_variables_idx state_variables_idx mf0 mf1
persistent sample_size number_of_state_variables number_of_observed_variables number_of_structural_innovations
% Set defaults.
if (nargin<4) || (nargin==4 && isempty(number_of_particles))
number_of_particles = 1000 ;
end
if nargin==2 || isempty(start)
start = 1;
end
dr = reduced_form_model.state.dr;% Decision rules and transition equations.
Q = reduced_form_model.state.Q;% Covariance matrix of the structural innovations.
H = reduced_form_model.measurement.H;% Covariance matrix of the measurement errors.
% Set persistent variables.
if isempty(init_flag)
mf0 = bayestopt_.mf0;
mf1 = bayestopt_.mf1;
restrict_variables_idx = bayestopt_.restrict_var_list;
observed_variables_idx = restrict_variables_idx(mf1);
state_variables_idx = restrict_variables_idx(mf0);
sample_size = size(Y,2);
number_of_state_variables = length(mf0);
number_of_observed_variables = length(mf1);
number_of_structural_innovations = length(Q);
init_flag = 1;
end
% Set local state space model (second order approximation).
ghx = dr.ghx(restrict_variables_idx,:);
ghu = dr.ghu(restrict_variables_idx,:);
half_ghxx = .5*dr.ghxx(restrict_variables_idx,:);
half_ghuu = .5*dr.ghuu(restrict_variables_idx,:);
ghxu = dr.ghxu(restrict_variables_idx,:);
steadystate = dr.ys(dr.order_var(restrict_variables_idx));
constant = steadystate + .5*dr.ghs2(restrict_variables_idx);
state_variables_steady_state = dr.ys(dr.order_var(state_variables_idx));
StateVectorMean = state_variables_steady_state;
StateVectorVariance = lyapunov_symm(ghx(mf0,:),ghu(mf0,:)*Q*ghu(mf0,:)',1e-12,1e-12);
StateVectorVarianceSquareRoot = reduced_rank_cholesky(StateVectorVariance)';
state_variance_rank = size(StateVectorVarianceSquareRoot,2);
%state_idx = 1:state_variance_rank;
%innovation_idx = 1+state_variance_rank:state_variance_rank+number_of_structural_innovations;
Q_lower_triangular_cholesky = chol(Q)';
% Set seed for randn().
seed = [ 362436069 ; 521288629 ];
randn('state',seed);
const_lik = log(2*pi)*number_of_observed_variables;
lik = NaN(sample_size,1);
nb_obs_resamp = 0 ;
for t=1:sample_size
PredictedState = zeros(number_of_particles,number_of_state_variables);
PredictionError = zeros(number_of_particles,number_of_observed_variables);
%PredictedStateMean = zeros(number_of_state_variables,1);
PredictedObservedMean = zeros(number_of_observed_variables,1);
%PredictedStateVariance = zeros(number_of_state_variables,number_of_state_variables);
PredictedObservedVariance = zeros(number_of_observed_variables,number_of_observed_variables);
%PredictedStateAndObservedCovariance = zeros(number_of_state_variables,number_of_observed_variables);
for i=1:number_of_particles
if t==1
StateVector = StateVectorMean + StateVectorVarianceSquareRoot*randn(state_variance_rank,1);
else
StateVector = StateUpdated(i,:)' ;
end
yhat = StateVector-state_variables_steady_state;
epsilon = Q_lower_triangular_cholesky*randn(number_of_structural_innovations,1);
tmp = local_state_space_iteration_2(yhat,epsilon,ghx,ghu,constant,half_ghxx,half_ghuu,ghxu);
% stockage des particules et des erreurs de prvisions
PredictedState(i,:) = tmp(mf0)' ;
PredictionError(i,:) = (Y(:,t) - tmp(mf1))' ;
% calcul des moyennes et des matrices de variances covariances
%PredictedStateMean_old = PredictedStateMean;
PredictedObservedMean_old = PredictedObservedMean;
%PredictedStateMean = PredictedStateMean + (tmp(mf0)-PredictedStateMean)/i;
PredictedObservedMean = PredictedObservedMean + (tmp(mf1)-PredictedObservedMean)/i;
%psm = PredictedStateMean*PredictedStateMean';
pom = PredictedObservedMean*PredictedObservedMean';
%pcm = PredictedStateMean*PredictedObservedMean';
%PredictedStateVariance = PredictedStateVariance ...
% + ( (tmp(mf0)*tmp(mf0)'-psm-PredictedStateVariance)+(i-1)*(PredictedStateMean_old*PredictedStateMean_old'-psm) )/i;
PredictedObservedVariance = PredictedObservedVariance ...
+ ( (tmp(mf1)*tmp(mf1)'-pom-PredictedObservedVariance)+(i-1)*(PredictedObservedMean_old*PredictedObservedMean_old'-pom) )/i;
%PredictedStateAndObservedCovariance = PredictedStateAndObservedCovariance ...
% + ( (tmp(mf0)*tmp(mf1)'-pcm-PredictedStateAndObservedCovariance)+(i-1)*(PredictedStateMean_old*PredictedObservedMean_old'-pcm) )/i;
end
PredictedObservedVariance = PredictedObservedVariance + H;
iPredictedObservedVariance = inv(PredictedObservedVariance);
lnw = - 0.5*(const_lik + log(det(PredictedObservedVariance)) + sum(PredictionError'*iPredictedObservedVariance.*PredictionError',2)) ;
%bidouille numrique Schorfheide
dfac = max(lnw);
wtilde = w.*exp(lnw - dfac) ;
% vraisemblance de l'observation
lik(t) = log(mean(wtilde)) + dfac ;
clear (PredictionError) ;
clear (lnw) ;
% calcul des poids
w = wtilde/sum(wtilde) ;
clear (wtilde) ;
%update
Neff = 1/sum(w^2) ;
if Neff>number_of_particules %no resampling
StateUpdated = PredictedState ;
clear (PredictedState) ;
w = number_of_particles*w ;
else %resampling
nb_obs_resamp = nb_obs_resamp+1 ;
%kill the smallest particles before resampling :! facultatif ?
to_kill = [w PredictedState] ;
to_kill = delif(to_kill,w<(1/number_of_particules)*1E-12);%%
[n,m] = size(to_kill) ;
w = to_kill(:,1) ;
PredictedState = to_kill(:,2:m) ;
clear (to_kill) ;
if number_of_particles neq n
'Elimination de '; number_of_particles - n ; ' particules l''observation ';t ;
end
%fin de kill
%remise l'chelle des poids sur les particules restantes
w = cumsum( w/sum(w) );
%Rchantillonage systmatique
rnduvec = ( (1:number_of_particles)-1+rand )/number_of_particles ;
selind = (n - sum( w > rnduvec' ) + 1)'; % problme de mmoire car w .> rnduvec' trs grande !
clear (rnduvec) ;
StateUpdated = PredictedState(selind,:) ;
clear (selind) ;
% initialize
selind = zeros(1,number_of_particles);
% construct CDF
c = cumsum(w);
% draw a starting point
rnduvec = ( (1:number_of_particles)-1+rand)/number_of_particles ;
% start at the bottom of the CDF
j=1;
for i=1:number_of_particles
% move along the CDF
while (rnduvec(i)>c(j))
j=j+1;
end
% assign index
selind(i) = j;
end
StateUpdated = PredictedState(selind,:);
w = ones(number_of_particules,1) ;
end
end
LIK = -sum(lik(start:end));
\ No newline at end of file
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment