| 1 |
function [flag] = compare_kalman_mex(experience) |
| 2 |
|
| 3 |
|
| 4 |
rng(experience.Seed); |
| 5 |
|
| 6 |
N = experience.NumberOfSimulations; |
| 7 |
|
| 8 |
pp = experience.Number0fObservedVariables*experience.Scale; |
| 9 |
mm = experience.SizeOfTheStateVector*experience.Scale; |
| 10 |
rr = experience.NumberOfStructuralShocks*experience.Scale; |
| 11 |
measurement_error_flag = experience.MeasurementErrors; |
| 12 |
gend = experience.NumberOfPeriods; |
| 13 |
|
| 14 |
%% SET VARIOUS PARAMETERS |
| 15 |
kalman_tol = 1e-12; |
| 16 |
riccati_tol =1e-9; |
| 17 |
|
| 18 |
%% SET THE STATE SPACE MODEL: |
| 19 |
|
| 20 |
% I randomly choose the mm eigenvalues of the transition matrix. |
| 21 |
TransitionEigenvalues = rand(mm, 1)*2-1; |
| 22 |
|
| 23 |
% I randomly choose the mm eigenvectors of the transition matrix |
| 24 |
TransitionEigenvectors = rand(mm, mm); |
| 25 |
|
| 26 |
% I build the transition matrix |
| 27 |
T = TransitionEigenvectors*diag(TransitionEigenvalues)/TransitionEigenvectors; |
| 28 |
|
| 29 |
% I randomly choose matrix R |
| 30 |
R = randn(mm, rr); |
| 31 |
|
| 32 |
% I randomly choose the covariance matrix of the structural innovations |
| 33 |
E = randn(rr, 20*rr); |
| 34 |
Q = E*transpose(E)/(20*rr); |
| 35 |
|
| 36 |
% If needed I randomly choose the covariance matrix of the measurement errors |
| 37 |
if measurement_error_flag == 0 |
| 38 |
H = zeros(pp,1); |
| 39 |
elseif measurement_error_flag == 1 |
| 40 |
H = rand(pp,1); |
| 41 |
elseif measurement_error_flag == 2 |
| 42 |
E = randn(pp,20*pp); |
| 43 |
H = E*transpose(E)/(20*pp); |
| 44 |
H = (.1*eye(pp))*H*(.1*eye(pp)); |
| 45 |
else |
| 46 |
disp('compare_kalman_mex: unknown option!') |
| 47 |
end |
| 48 |
|
| 49 |
% Set the selection vector (mf) |
| 50 |
MF = transpose(randperm(mm)); |
| 51 |
mf = MF(1:pp); |
| 52 |
|
| 53 |
% Compute ergodic variance of the state equation |
| 54 |
P = lyapunov_symm(T,R*Q*R',riccati_tol,1.000001, riccati_tol); |
| 55 |
|
| 56 |
if measurement_error_flag==0 |
| 57 |
HH = []; |
| 58 |
elseif measurement_error_flag==1 |
| 59 |
HH = diag(H); |
| 60 |
elseif measurement_error_flag==2 |
| 61 |
HH = H; |
| 62 |
end |
| 63 |
|
| 64 |
rng(experience.Seed*1938); |
| 65 |
|
| 66 |
% Build datasets |
| 67 |
Y = simul_state_space_model(T,R,Q,mf,gend*N, HH); |
| 68 |
|
| 69 |
if isempty(HH) |
| 70 |
HH = 0; |
| 71 |
end |
| 72 |
|
| 73 |
% |
| 74 |
% Evaluate likelihoods |
| 75 |
% |
| 76 |
|
| 77 |
LIK_matlab_0 = NaN(N,1); |
| 78 |
LIK_mex_0 = NaN(N,1); |
| 79 |
|
| 80 |
flag = 0; |
| 81 |
Zflag = 0; |
| 82 |
tic; |
| 83 |
for i=1:N |
| 84 |
[LIK_matlab_0(i), ~] = kalman_filter(Y(:,(i-1)*gend+1:i*gend),1,gend,zeros(mm,1),P,kalman_tol,riccati_tol,0,0,T,Q,R,HH,mf,mm,pp,rr,Zflag,0,0); |
| 85 |
end |
| 86 |
T_matlab_0 = toc; |
| 87 |
|
| 88 |
tic; |
| 89 |
for i=1:N |
| 90 |
[LIK_mex_0(i), ~] = kalman_filter_mex(Y(:,(i-1)*gend+1:i*gend),zeros(mm,1),P,kalman_tol,riccati_tol,T,Q,R,mf,Zflag,HH); |
| 91 |
end |
| 92 |
T_mex_0 = toc; |
| 93 |
|
| 94 |
dprintf('Zflag = 0: Fortran Kalman filter is %5.2f times faster than its Matlab counterpart.', T_matlab_0/T_mex_0) |
| 95 |
|
| 96 |
if max(abs((LIK_mex_0-LIK_matlab_0)./LIK_matlab_0))>1e-6 |
| 97 |
dprintf("Zflag = 0: discrepancy between Matlab and Fortran Kalman filter results!") |
| 98 |
flag = 1; |
| 99 |
end |
| 100 |
|
| 101 |
LIK_matlab_1 = NaN(N,1); |
| 102 |
LIK_mex_1 = NaN(N,1); |
| 103 |
|
| 104 |
Zflag = 1; |
| 105 |
Z = eye(mm); |
| 106 |
Z = Z(mf,:); |
| 107 |
tic; |
| 108 |
for i=1:N |
| 109 |
[LIK_matlab_1(i), ~] = kalman_filter(Y(:,(i-1)*gend+1:i*gend),1,gend,zeros(mm,1),P,kalman_tol,riccati_tol,0,0,T,Q,R,HH,Z,mm,pp,rr,Zflag,0,0); |
| 110 |
end |
| 111 |
T_matlab_1 = toc; |
| 112 |
|
| 113 |
tic; |
| 114 |
for i=1:N |
| 115 |
[LIK_mex_1(i), ~] = kalman_filter_mex(Y(:,(i-1)*gend+1:i*gend),zeros(mm,1),P,kalman_tol,riccati_tol,T,Q,R,Z,Zflag,HH); |
| 116 |
end |
| 117 |
T_mex_1 = toc; |
| 118 |
|
| 119 |
dprintf('Zflag = 1: Fortran Kalman filter is %5.2f times faster than its Matlab counterpart.', T_matlab_1/T_mex_1) |
| 120 |
|
| 121 |
if max(abs((LIK_mex_1-LIK_matlab_1)./LIK_matlab_1))>1e-6 |
| 122 |
dprintf("Zflag = 1: discrepancy between Matlab and Fortran Kalman filter results!") |
| 123 |
flag = 1; |
| 124 |
end |
| 125 |
|
| 126 |
if max(abs((LIK_mex_0 - LIK_mex_1)./LIK_mex_1))>1e-6 |
| 127 |
dprintf("Zflag = 1: discrepancy between results with and without the Zflag!") |
| 128 |
flag = 1; |
| 129 |
end |
| 130 |
|
| 131 |
dprintf('Cost of Zflag==1 is %5.2f%% (matlab)', 100*(T_matlab_1-T_matlab_0)/T_matlab_0) |
| 132 |
dprintf('Cost of Zflag==1 is %5.2f%% (mex)', 100*(T_mex_1-T_mex_0)/T_mex_0) |