1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
|
function [fhat,xhat,fcount,retcode] = csminit1(fcn,x0,penalty,f0,g0,badg,H0,Verbose,varargin)
% [fhat,xhat,fcount,retcode] = csminit1(fcn,x0,penalty,f0,g0,badg,H0,Verbose,varargin)
%
% Inputs:
% fcn: [string] string naming the objective function to be minimized
% x0: [npar by 1] initial value of the parameter vector
% penalty: [scalar] variable penalty in case of failure of objective function
% f0: [scalar] initial value of the fucntion
% g0: [npar by 1] initial value of the gradient vector
% badg [scalar] indicator for problem in gradient computation
% H0: [npar by npar] initial value for the inverse Hessian. Must be positive definite.
% Verbose: [scalar] indicator for verbose computations
% varargin: Optional additional inputs that get handed off to fcn each
% time it is called.
% Outputs:
% fhat: [scalar] function value at minimum
% xhat: [npar by 1] parameter vector at minimum
% fcount [scalar] function iteration count upon termination
% retcode [scalar] 0: normal step
% 1: zero gradient.
% 5: largest step still improves too fast.
% 2,4: back and forth adjustment of stepsize didn't finish.
% 3: smallest stepsize still improves too slow
% 6: no improvement found
%---------------------
% Modified 7/22/96 to omit variable-length P list, for efficiency and compilation.
% Places where the number of P's need to be altered or the code could be returned to
% its old form are marked with ARGLIST comments.
%
% Fixed 7/17/93 to use inverse-hessian instead of hessian itself in bfgs
% update.
%
% Fixed 7/19/93 to flip eigenvalues of H to get better performance when
% it's not psd.
%
% Original file downloaded from:
% http://sims.princeton.edu/yftp/optimize/mfiles/csminit.m
%
% Copyright (C) 1993-2007 Christopher Sims
% Copyright (C) 2008-2017 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/>.
%tailstr = ')';
%for i=nargin-6:-1:1
% tailstr=[ ',P' num2str(i) tailstr];
%end
%ANGLE = .03;
ANGLE = .005;
%THETA = .03;
THETA = .3; %(0<THETA<.5) THETA near .5 makes long line searches, possibly fewer iterations.
FCHANGE = 1000;
MINLAMB = 1e-9;
% fixed 7/15/94
% MINDX = .0001;
% MINDX = 1e-6;
MINDFAC = .01;
fcount=0;
lambda=1;
xhat=x0;
f=f0;
fhat=f0;
g = g0;
gnorm = norm(g);
%
if (gnorm < 1.e-12) && ~badg % put ~badg 8/4/94
retcode =1;
dxnorm=0;
% gradient convergence
else
% with badg true, we don't try to match rate of improvement to directional
% derivative. We're satisfied just to get some improvement in f.
%
%if(badg)
% dx = -g*FCHANGE/(gnorm*gnorm);
% dxnorm = norm(dx);
% if dxnorm > 1e12
% disp('Bad, small gradient problem.')
% dx = dx*FCHANGE/dxnorm;
% end
%else
% Gauss-Newton step;
%---------- Start of 7/19/93 mod ---------------
%[v d] = eig(H0);
%toc
%d=max(1e-10,abs(diag(d)));
%d=abs(diag(d));
%dx = -(v.*(ones(size(v,1),1)*d'))*(v'*g);
% toc
dx = -H0*g;
% toc
dxnorm = norm(dx);
if dxnorm > 1e12
disp_verbose('Near-singular H problem.',Verbose)
dx = dx*FCHANGE/dxnorm;
end
dfhat = dx'*g0;
%end
%
%
if ~badg
% test for alignment of dx with gradient and fix if necessary
a = -dfhat/(gnorm*dxnorm);
if a<ANGLE
dx = dx - (ANGLE*dxnorm/gnorm+dfhat/(gnorm*gnorm))*g;
dfhat = dx'*g;
dxnorm = norm(dx);
disp_verbose(sprintf('Correct for low angle: %g',a),Verbose)
end
end
disp_verbose(sprintf('Predicted improvement: %18.9f',-dfhat/2),Verbose)
%
% Have OK dx, now adjust length of step (lambda) until min and
% max improvement rate criteria are met.
done=0;
factor=3;
shrink=1;
lambdaMin=0;
lambdaMax=inf;
lambdaPeak=0;
fPeak=f0;
lambdahat=0;
while ~done
if size(x0,2)>1
dxtest=x0+dx'*lambda;
else
dxtest=x0+dx*lambda;
end
% home
f = penalty_objective_function(dxtest,fcn,penalty,varargin{:});
%ARGLIST
%f = feval(fcn,dxtest,P1,P2,P3,P4,P5,P6,P7,P8,P9,P10,P11,P12,P13);
% f = feval(fcn,x0+dx*lambda,P1,P2,P3,P4,P5,P6,P7,P8);
disp_verbose(sprintf('lambda = %10.5g; f = %20.7f',lambda,f ),Verbose)
%debug
%disp(sprintf('Improvement too great? f0-f: %g, criterion: %g',f0-f,-(1-THETA)*dfhat*lambda))
if f<fhat
fhat=f;
xhat=dxtest;
lambdahat = lambda;
end
fcount=fcount+1;
shrinkSignal = (~badg & (f0-f < max([-THETA*dfhat*lambda 0]))) | (badg & (f0-f) < 0) ;
growSignal = ~badg & ( (lambda > 0) & (f0-f > -(1-THETA)*dfhat*lambda) );
if shrinkSignal && ( (lambda>lambdaPeak) || (lambda<0) )
if (lambda>0) && ((~shrink) || (lambda/factor <= lambdaPeak))
shrink=1;
factor=factor^.6;
while lambda/factor <= lambdaPeak
factor=factor^.6;
end
%if (abs(lambda)*(factor-1)*dxnorm < MINDX) || (abs(lambda)*(factor-1) < MINLAMB)
if abs(factor-1)<MINDFAC
if abs(lambda)<4
retcode=2;
else
retcode=7;
end
done=1;
end
end
if (lambda<lambdaMax) && (lambda>lambdaPeak)
lambdaMax=lambda;
end
lambda=lambda/factor;
if abs(lambda) < MINLAMB
if (lambda > 0) && (f0 <= fhat)
% try going against gradient, which may be inaccurate
if Verbose
lambda = -lambda*factor^6
else
lambda = -lambda*factor^6;
end
else
if lambda < 0
retcode = 6;
else
retcode = 3;
end
done = 1;
end
end
elseif (growSignal && lambda>0) || (shrinkSignal && ((lambda <= lambdaPeak) && (lambda>0)))
if shrink
shrink=0;
factor = factor^.6;
%if ( abs(lambda)*(factor-1)*dxnorm< MINDX ) || ( abs(lambda)*(factor-1)< MINLAMB)
if abs(factor-1)<MINDFAC
if abs(lambda)<4
retcode=4;
else
retcode=7;
end
done=1;
end
end
if ( f<fPeak ) && (lambda>0)
fPeak=f;
lambdaPeak=lambda;
if lambdaMax<=lambdaPeak
lambdaMax=lambdaPeak*factor*factor;
end
end
lambda=lambda*factor;
if abs(lambda) > 1e20
retcode = 5;
done =1;
end
else
done=1;
if factor < 1.2
retcode=7;
else
retcode=0;
end
end
end
end
disp_verbose(sprintf('Norm of dx %10.5g', dxnorm),Verbose)
|
