Commit 0011678c authored by sebastien's avatar sebastien
Browse files

Changes by Ferhat:

* fix options stack_solve_algo={2,3,4} (closes #68)
* fix crashes for singular normalizations (closes #44) and implement decreasing cutoff
* fail for stack_solve_algo=2 under Octave (because there is no gmres function in Octave)


git-svn-id: https://www.dynare.org/svn/dynare/trunk@3279 ac1d8469-bf42-47a9-8791-bf33cf982152
parent 3885bddd
......@@ -1836,7 +1836,7 @@ steady(homotopy_mode = 1, homotopy_steps = 50);
<itemizedlist>
<listitem><para><literal>0</literal>: Newton method to solve simultaneously all the equations for every period, see <xref linkend="juillard_1996"/>. (Default)</para></listitem>
<listitem><para><literal>1</literal>: use a Newton algorithm with a sparse LU solver at each iteration.</para></listitem>
<listitem><para><literal>2</literal>: use a Newton algorithm with a Generalized Minimal Residual (GMRES) solver at each iteration.</para></listitem>
<listitem><para><literal>2</literal>: use a Newton algorithm with a Generalized Minimal Residual (GMRES) solver at each iteration. This option is not available under Octave.</para></listitem>
<listitem><para><literal>3</literal>: use a Newton algorithm with a Stabilized Bi-Conjugate Gradient (BICGSTAB) solver at each iteration.</para></listitem>
<listitem><para><literal>4</literal>: use a Newton algorithm with a optimal path length at each iteration.</para></listitem>
<listitem><para><literal>5</literal>: use a Newton algorithm with a sparse Gaussian elimination (SPE) solver at each iteration.</para></listitem>
......
function r = lnsrch1_wrapper_one_boundary(ya, y_index, fname, y, x, params, it_)
% wrapper for solve_one_boundary m-file when it is used with a dynamic
% model
%
% INPUTS
% ya [vector] The endogenous of the current block
% y_index [vector of int] The index of the endogenous variables of
% the block
% fname [string] name of the file containing the block
% to simulate
% y [matrix] All the endogenous variables of the model
% x [matrix] All the exogenous variables of the model
% params [vector] All the parameters of the model
% OUTPUTS
% r [vector] The residuals of the current block
%
% ALGORITHM
% none.
%
% SPECIAL REQUIREMENTS
% none.
%
% 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/licen
%reshape the input arguments of the dynamic function
y(it_, :) = ya;
[r, y, g1, g2, g3]=feval(fname, y, x, params, it_, 0);
function ra = lnsrch1_wrapper_two_boundaries(ya, fname, y, y_index, x, params, periods, y_kmin, y_size)
% wrapper for solve_one_boundary m-file when it is used with a dynamic
% model
%
% INPUTS
% ya [vector] The endogenous of the current block
% y_index [vector of int] The index of the endogenous variables of
% the block
% fname [string] name of the file containing the block
% to simulate
% y [matrix] All the endogenous variables of the model
% x [matrix] All the exogenous variables of the model
% params [vector] All the parameters of the model
% periods [int] The number of periods
% y_kmin [int] The maximum number of lag on en endogenous variables
% y_size [int] The number of endogenous variables
% in the current block
% OUTPUTS
% ra [vector] The residuals of the current block
%
% ALGORITHM
% none.
%
% SPECIAL REQUIREMENTS
% none.
%
% 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/licen
%reshape the input arguments of the dynamic function
y(y_kmin+1:y_kmin+periods, y_index) = reshape(ya',length(y_index),periods)';
[r, y, g1, g2, g3, b]=feval(fname, y, x, params, periods, 0, y_kmin, y_size);
ra = reshape(r(:, y_kmin+1:periods+y_kmin),periods*y_size, 1);
......@@ -67,6 +67,10 @@ if options_.block && options_.bytecode && options_.stack_solve_algo ~= 5
error('SIMUL: for the moment, you must use stack_solve_algo=5 with block and bytecode option')
end
if exist('OCTAVE_VERSION') && options_.stack_solve_algo == 2
error('SIMUL: stack_solve_algo=2 is not available for Octave. Choose another value.')
end
if(options_.block)
if(options_.bytecode)
oo_.endo_simul=bytecode('dynamic');
......
......@@ -237,20 +237,25 @@ for it_=start:incr:finish
lambda=1;
stpmx = 100 ;
if (is_dynamic)
stpmax = stpmx*max([sqrt(y*y');size(y_index_eq,2)]);
stpmax = stpmx*max([sqrt(ya'*ya);size(y_index_eq,2)]);
else
stpmax = stpmx*max([sqrt(y'*y);size(y_index_eq,2)]);
stpmax = stpmx*max([sqrt(ya'*ya);size(y_index_eq,2)]);
end;
nn=1:size(y_index_eq,2);
g = (r'*g1)';
f = 0.5*r'*r;
p = -g1\r ;
if (is_dynamic)
[y,f,r,check]=lnsrch1(y,f,g,p,stpmax,fname,nn,y_index_eq,x, params, it_, 0);
[ya,f,r,check]=lnsrch1(y(it_,:),f,g,p,stpmax,'lnsrch1_wrapper_one_boundary',nn, y_index_eq, y_index_eq, fname, y, x, params, it_);
else
[y,f,r,check]=lnsrch1(y,f,g,p,stpmax,fname,nn,y_index_eq,x, params, 0);
[ya,f,r,check]=lnsrch1(y,f,g,p,stpmax,fname,nn,y_index_eq,x, params, 0);
end;
dx = ya - y(y_index_eq);
if(is_dynamic)
y(it_,:) = ya';
else
y = ya';
end;
elseif(~is_dynamic & options_.solve_algo==3)
[yn,info] = csolve(@local_fname, y(y_index_eq),@local_fname,1e-6,500, x, params, y, y_index_eq, fname, 1);
dx = ya - yn;
......@@ -267,8 +272,7 @@ for it_=start:incr:finish
flag1=1;
while(flag1>0)
[L1, U1]=luinc(g1,luinc_tol);
[za,flag1] = gmres(g1,-r,Blck_size,1e-6,Blck_size,L1,U1);
%[za,flag1] = gmres(-g1,b',Blck_size,1e-6,Blck_size,L1,U1);
[dx,flag1] = gmres(g1,-r,Blck_size,1e-6,Blck_size,L1,U1);
if (flag1>0 | reduced)
if(flag1==1)
disp(['Error in simul: No convergence inside GMRES after ' num2str(iter,'%6d') ' iterations, in block' num2str(Block_Num,'%3d')]);
......@@ -280,7 +284,6 @@ for it_=start:incr:finish
luinc_tol = luinc_tol/10;
reduced = 0;
else
dx = za - ya;
ya = ya + lambda*dx;
if(is_dynamic)
y(it_,y_index_eq) = ya';
......@@ -293,8 +296,7 @@ for it_=start:incr:finish
flag1=1;
while(flag1>0)
[L1, U1]=luinc(g1,luinc_tol);
[za,flag1] = bicgstab(g1,-r,1e-7,Blck_size,L1,U1);
%[za,flag1] = bicgstab(-g1,b',1e-7,Blck_size,L1,U1);
[dx,flag1] = bicgstab(g1,-r,1e-7,Blck_size,L1,U1);
if (flag1>0 | reduced)
if(flag1==1)
disp(['Error in simul: No convergence inside BICGSTAB after ' num2str(iter,'%6d') ' iterations, in block' num2str(Block_Num,'%3d')]);
......@@ -306,7 +308,6 @@ for it_=start:incr:finish
luinc_tol = luinc_tol/10;
reduced = 0;
else
dx = za - ya;
ya = ya + lambda*dx;
if(is_dynamic)
y(it_,y_index_eq) = ya';
......
......@@ -213,14 +213,6 @@ while ~(cvg==1 | iter>maxit_),
dx = g1a\b- ya;
ya = ya + lambda*dx;
y(1+y_kmin:periods+y_kmin,y_index)=reshape(ya',length(y_index),periods)';
% v = '';
% for i=1:(size(y_index,2))
% v = [v ' %1.6f'];
% end;
% v = [v '\n'];
% v
% sprintf(v,y(:,y_index)')
% return;
elseif(stack_solve_algo==2),
flag1=1;
while(flag1>0)
......@@ -264,11 +256,20 @@ while ~(cvg==1 | iter>maxit_),
end;
end;
elseif(stack_solve_algo==4),
error('SOLVE_TWO_BOUNDARIES: stack_solve_algo=4 not implemented')
end;
ra = reshape(r(:, y_kmin+1:periods+y_kmin),periods*Blck_size, 1);
stpmx = 100 ;
stpmax = stpmx*max([sqrt(ya'*ya);size(y_index,2)]);
nn=1:size(ra,1);
g = (ra'*g1a)';
f = 0.5*ra'*ra;
p = -g1a\ra;
[yn,f,ra,check]=lnsrch1(ya,f,g,p,stpmax,'lnsrch1_wrapper_two_boundaries',nn,nn, fname, y, y_index, x, params, periods, y_kmin, Blck_size);
dx = ya - yn;
y(1+y_kmin:periods+y_kmin,y_index)=reshape(yn',length(y_index),periods)';
end
end
iter=iter+1;
disp(['iteration: ' num2str(iter,'%d') ' error: ' num2str(max_res,'%e')]);
disp(['iteration: ' num2str(iter,'%d') ' error: ' num2str(max_res,'%e') ' stack_solve_algo=' num2str(stack_solve_algo)]);
end;
if (iter>maxit_)
disp(['No convergence after ' num2str(iter,'%4d') ' iterations in Block ' num2str(Block_Num,'%d')]);
......
......@@ -1995,7 +1995,7 @@ DynamicModel::computingPass(bool jacobianExo, bool hessian, bool thirdDerivative
evaluateAndReduceJacobian(eval_context, contemporaneous_jacobian, static_jacobian, dynamic_jacobian, cutoff, false);
computePossiblySingularNormalization(contemporaneous_jacobian, cutoff == 0);
computeNonSingularNormalization(contemporaneous_jacobian, cutoff, static_jacobian, dynamic_jacobian);
computePrologueAndEpilogue(static_jacobian, equation_reordered, variable_reordered, prologue, epilogue);
......
......@@ -32,8 +32,8 @@
using namespace boost;
using namespace MFS;
void
ModelTree::computeNormalization(const set<pair<int, int> > &endo_eqs_incidence) throw (NormalizationException)
bool
ModelTree::computeNormalization(const jacob_map &contemporaneous_jacobian, bool verbose)
{
const int n = equation_number();
......@@ -50,8 +50,8 @@ ModelTree::computeNormalization(const set<pair<int, int> > &endo_eqs_incidence)
// Fill in the graph
set<pair<int, int> > endo;
for (set<pair<int, int> >::const_iterator it = endo_eqs_incidence.begin(); it != endo_eqs_incidence.end(); it++)
add_edge(it->first + n, it->second, g);
for (jacob_map::const_iterator it = contemporaneous_jacobian.begin(); it != contemporaneous_jacobian.end(); it++)
add_edge(it->first.first + n, it->first.second, g);
// Compute maximum cardinality matching
vector<int> mate_map(2*n);
......@@ -125,63 +125,96 @@ ModelTree::computeNormalization(const set<pair<int, int> > &endo_eqs_incidence)
// Check if all variables are normalized
vector<int>::const_iterator it = find(mate_map.begin(), mate_map.begin() + n, graph_traits<BipartiteGraph>::null_vertex());
if (it != mate_map.begin() + n)
throw NormalizationException(symbol_table.getID(eEndogenous, it - mate_map.begin()));
{
if (verbose)
cerr << "ERROR: Could not normalize the model. Variable "
<< symbol_table.getName(symbol_table.getID(eEndogenous, it - mate_map.begin()))
<< " is not in the maximum cardinality matching." << endl;
check = false;
}
return check;
}
void
ModelTree::computePossiblySingularNormalization(const jacob_map &contemporaneous_jacobian, bool try_symbolic)
ModelTree::computeNonSingularNormalization(jacob_map &contemporaneous_jacobian, double cutoff, jacob_map &static_jacobian, dynamic_jacob_map &dynamic_jacobian)
{
bool check = false;
cout << "Normalizing the model..." << endl;
set<pair<int, int> > endo_eqs_incidence;
int n = equation_number();
for (jacob_map::const_iterator it = contemporaneous_jacobian.begin();
it != contemporaneous_jacobian.end(); it++)
endo_eqs_incidence.insert(make_pair(it->first.first, it->first.second));
// compute the maximum value of each row of the contemporaneous Jacobian matrix
//jacob_map normalized_contemporaneous_jacobian;
jacob_map normalized_contemporaneous_jacobian(contemporaneous_jacobian);
vector<double> max_val(n, 0.0);
for (jacob_map::const_iterator iter = contemporaneous_jacobian.begin(); iter != contemporaneous_jacobian.end(); iter++)
if (fabs(iter->second) > max_val[iter->first.first])
max_val[iter->first.first] = fabs(iter->second);
try
{
computeNormalization(endo_eqs_incidence);
return;
}
catch (NormalizationException &e)
for (jacob_map::iterator iter = normalized_contemporaneous_jacobian.begin(); iter != normalized_contemporaneous_jacobian.end(); iter++)
iter->second /= max_val[iter->first.first];
//We start with the highest value of the cutoff and try to normalize the model
double current_cutoff = 0.99999999;
int suppressed = 0;
while (!check && current_cutoff > 1e-19)
{
if (try_symbolic)
cout << "Normalization failed with cutoff, trying symbolic normalization..." << endl;
else
jacob_map tmp_normalized_contemporaneous_jacobian;
int suppress = 0;
for (jacob_map::iterator iter = normalized_contemporaneous_jacobian.begin(); iter != normalized_contemporaneous_jacobian.end(); iter++)
if (fabs(iter->second) > max(current_cutoff, cutoff))
tmp_normalized_contemporaneous_jacobian[make_pair(iter->first.first, iter->first.second)] = iter->second;
else
suppress++;
if (suppress != suppressed)
check = computeNormalization(tmp_normalized_contemporaneous_jacobian, false);
suppressed = suppress;
if (!check)
{
cerr << "ERROR: Could not normalize the model. Variable "
<< symbol_table.getName(e.symb_id)
<< " is not in the maximum cardinality matching. Try to decrease the cutoff." << endl;
exit(EXIT_FAILURE);
current_cutoff /= 2;
// In this last case try to normalize with the complete jacobian
if (current_cutoff <= 1e-19)
check = computeNormalization(normalized_contemporaneous_jacobian, false);
}
}
// If no non-singular normalization can be found, try to find a normalization even with a potential singularity
if (try_symbolic)
if (!check)
{
endo_eqs_incidence.clear();
cout << "Normalization failed with cutoff, trying symbolic normalization..." << endl;
//if no non-singular normalization can be found, try to find a normalization even with a potential singularity
jacob_map tmp_normalized_contemporaneous_jacobian;
set<pair<int, int> > endo;
for (int i = 0; i < equation_number(); i++)
for (int i = 0; i < n; i++)
{
endo.clear();
equations[i]->collectEndogenous(endo);
for (set<pair<int, int> >::const_iterator it = endo.begin(); it != endo.end(); it++)
endo_eqs_incidence.insert(make_pair(i, it->first));
tmp_normalized_contemporaneous_jacobian[make_pair(i, it->first)] = 1;
}
try
check = computeNormalization(tmp_normalized_contemporaneous_jacobian, true);
if (check)
{
computeNormalization(endo_eqs_incidence);
}
catch (NormalizationException &e)
{
cerr << "ERROR: Could not normalize the model even with zero cutoff. Variable "
<< symbol_table.getName(e.symb_id)
<< " is not in the maximum cardinality matching." << endl;
exit(EXIT_FAILURE);
// Update the jacobian matrix
for (jacob_map::const_iterator it = tmp_normalized_contemporaneous_jacobian.begin(); it != tmp_normalized_contemporaneous_jacobian.end(); it++)
{
if (static_jacobian.find(make_pair(it->first.first, it->first.second)) == static_jacobian.end())
static_jacobian[make_pair(it->first.first, it->first.second)] = 0;
if (dynamic_jacobian.find(make_pair(0, make_pair(it->first.first, it->first.second))) == dynamic_jacobian.end())
dynamic_jacobian[make_pair(0, make_pair(it->first.first, it->first.second))] = 0;
if (contemporaneous_jacobian.find(make_pair(it->first.first, it->first.second)) == contemporaneous_jacobian.end())
contemporaneous_jacobian[make_pair(it->first.first, it->first.second)] = 0;
}
}
}
if (!check)
{
cerr << "No normalization could be computed. Aborting." << endl;
exit(EXIT_FAILURE);
}
}
void
......
......@@ -136,24 +136,21 @@ protected:
//! for each block contains pair< max_lag, max_lead>
t_lag_lead_vector block_lag_lead;
//! Exception thrown when normalization fails
class NormalizationException
{
public:
//! A variable missing from the maximum cardinal matching
int symb_id;
NormalizationException(int symb_id_arg) : symb_id(symb_id_arg)
{
}
};
//! Compute the matching between endogenous and variable using the jacobian contemporaneous_jacobian
/*! \param endo_eqs_incidence A set indicating which endogenous appear in which equation. First element of pairs is equation number, second is type specific endo ID */
void computeNormalization(const set<pair<int, int> > &endo_eqs_incidence) throw (NormalizationException);
/*!
\param contemporaneous_jacobian Jacobian used as an incidence matrix: all elements declared in the map (even if they are zero), are used as vertices of the incidence matrix
\return True if a complete normalization has been achieved
*/
bool computeNormalization(const jacob_map &contemporaneous_jacobian, bool verbose);
//! Try to compute the matching between endogenous and variable using a decreasing cutoff
/*! applied to the jacobian contemporaneous_jacobian and stop when a matching is found.*/
/*! if no matching is found with a cutoff close to zero an error message is printout */
void computePossiblySingularNormalization(const jacob_map &contemporaneous_jacobian, bool try_symbolic);
/*!
Applied to the jacobian contemporaneous_jacobian and stop when a matching is found.
If no matching is found using a strictly positive cutoff, then a zero cutoff is applied (i.e. use a symbolic normalization); in that case, the method adds zeros in the jacobian matrices to reflect all the edges in the symbolic incidence matrix.
If no matching is found with a zero cutoff close to zero an error message is printout.
*/
void computeNonSingularNormalization(jacob_map &contemporaneous_jacobian, double cutoff, jacob_map &static_jacobian, dynamic_jacob_map &dynamic_jacobian);
//! Try to normalized each unnormalized equation (matched endogenous variable only on the LHS)
void computeNormalizedEquations(multimap<int, int> &endo2eqs) const;
//! Evaluate the jacobian and suppress all the elements below the cutoff
......
......@@ -45,7 +45,7 @@ StaticModel::StaticModel(SymbolTable &symbol_table_arg,
}
void
StaticModel::compileDerivative(ofstream &code_file, int eq, int symb_id, int lag, map_idx_type &map_idx) const
StaticModel::compileDerivative(ofstream &code_file, int eq, int symb_id, map_idx_type &map_idx) const
{
first_derivatives_type::const_iterator it = first_derivatives.find(make_pair(eq, symbol_table.getID(eEndogenous, symb_id)));
if (it != first_derivatives.end())
......@@ -539,7 +539,7 @@ StaticModel::writeModelEquationsCodeOrdered(const string file_name, const string
{
case SOLVE_BACKWARD_SIMPLE:
case SOLVE_FORWARD_SIMPLE:
compileDerivative(code_file, getBlockEquationID(block, 0), getBlockVariableID(block, 0), 0, map_idx);
compileDerivative(code_file, getBlockEquationID(block, 0), getBlockVariableID(block, 0), map_idx);
{
FSTPG_ fstpg(0);
fstpg.write(code_file);
......@@ -727,7 +727,7 @@ StaticModel::computingPass(const eval_context_type &eval_context, bool no_tmp_te
evaluateAndReduceJacobian(eval_context, contemporaneous_jacobian, static_jacobian, dynamic_jacobian, cutoff, false);
computePossiblySingularNormalization(contemporaneous_jacobian, cutoff == 0);
computeNonSingularNormalization(contemporaneous_jacobian, cutoff, static_jacobian, dynamic_jacobian);
computePrologueAndEpilogue(static_jacobian, equation_reordered, variable_reordered, prologue, epilogue);
......
......@@ -74,7 +74,7 @@ private:
void computeTemporaryTermsOrdered();
//! Write derivative code of an equation w.r. to a variable
void compileDerivative(ofstream &code_file, int eq, int symb_id, int lag, map_idx_type &map_idx) const;
void compileDerivative(ofstream &code_file, int eq, int symb_id, map_idx_type &map_idx) const;
//! Write chain rule derivative code of an equation w.r. to a variable
void compileChainRuleDerivative(ofstream &code_file, int eq, int var, int lag, map_idx_type &map_idx) const;
......
......@@ -13,6 +13,8 @@ OCTAVE_MODS = \
predetermined_variables.mod \
block_bytecode/fs2000_simk.mod \
block_bytecode/fs2000_lu.mod \
block_bytecode/fs2000_bicgstab.mod \
block_bytecode/fs2000_optpath.mod \
block_bytecode/fs2000_bytecode.mod \
block_bytecode/ramst.mod \
block_bytecode/ireland.mod \
......@@ -48,7 +50,9 @@ MODS = $(OCTAVE_MODS) \
AIM/ls2003_2L0L.mod \
AIM/ls2003_2L0L_AIM.mod \
AIM/ls2003_2L2L.mod \
AIM/ls2003_2L2L_AIM.mod
AIM/ls2003_2L2L_AIM.mod \
block_bytecode/fs2000_gmres.mod \
block_bytecode/ramst_a.mod
EXTRA_DIST = $(MODS) \
run_test_octave.m \
......
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