Chaste Release::3.1
SimpleNewtonNonlinearSolver.cpp
00001 /*
00002 
00003 Copyright (c) 2005-2012, University of Oxford.
00004 All rights reserved.
00005 
00006 University of Oxford means the Chancellor, Masters and Scholars of the
00007 University of Oxford, having an administrative office at Wellington
00008 Square, Oxford OX1 2JD, UK.
00009 
00010 This file is part of Chaste.
00011 
00012 Redistribution and use in source and binary forms, with or without
00013 modification, are permitted provided that the following conditions are met:
00014  * Redistributions of source code must retain the above copyright notice,
00015    this list of conditions and the following disclaimer.
00016  * Redistributions in binary form must reproduce the above copyright notice,
00017    this list of conditions and the following disclaimer in the documentation
00018    and/or other materials provided with the distribution.
00019  * Neither the name of the University of Oxford nor the names of its
00020    contributors may be used to endorse or promote products derived from this
00021    software without specific prior written permission.
00022 
00023 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
00024 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
00025 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
00026 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
00027 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
00028 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
00029 GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
00030 HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
00031 LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
00032 OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
00033 
00034 */
00035 
00036 #include "SimpleNewtonNonlinearSolver.hpp"
00037 #include <iostream>
00038 #include <cassert>
00039 #include "Exception.hpp"
00040 
00041 SimpleNewtonNonlinearSolver::SimpleNewtonNonlinearSolver(double linearSolverRelativeTolerance)
00042         : mLinearSolverRelativeTolerance(linearSolverRelativeTolerance),
00043         mTolerance(1e-5),
00044         mWriteStats(false)
00045 {
00046     mTestDampingValues.push_back(-0.1);
00047     mTestDampingValues.push_back(0.05);
00048     for (unsigned i=1; i<=12; i++)
00049     {
00050         double val = double(i)/10;
00051         mTestDampingValues.push_back(val);
00052     }
00053 }
00054 
00055 SimpleNewtonNonlinearSolver::~SimpleNewtonNonlinearSolver()
00056 {}
00057 
00058 Vec SimpleNewtonNonlinearSolver::Solve(PetscErrorCode (*pComputeResidual)(SNES,Vec,Vec,void*),
00059                                        PetscErrorCode (*pComputeJacobian)(SNES,Vec,Mat*,Mat*,MatStructure*,void*),
00060                                        Vec initialGuess,
00061                                        unsigned fill,
00062                                        void* pContext)
00063 {
00064     PetscInt size;
00065     VecGetSize(initialGuess, &size);
00066 
00067     Vec current_solution;
00068     VecDuplicate(initialGuess, &current_solution);
00069     VecCopy(initialGuess, current_solution);
00070 
00071     LinearSystem linear_system(current_solution, fill);
00072 
00073     (*pComputeResidual)(NULL, current_solution, linear_system.rGetRhsVector(), pContext);
00074 
00075 
00076     double residual_norm;
00077     VecNorm(linear_system.rGetRhsVector(), NORM_2, &residual_norm);
00078     double scaled_residual_norm = residual_norm/size;
00079 
00080     if (mWriteStats)
00081     {
00082         std::cout << "Newton's method:\n  Initial ||residual||/N = " << scaled_residual_norm
00083                   << "\n  Attempting to solve to tolerance " << mTolerance << "..\n";
00084     }
00085 
00086     double old_scaled_residual_norm;
00087     unsigned counter = 0;
00088     while (scaled_residual_norm > mTolerance)
00089     {
00090         counter++;
00091 
00092         // Store the old norm to check with the new later
00093         old_scaled_residual_norm = scaled_residual_norm;
00094 
00095         // Compute Jacobian and solve J dx = f for the (negative) update dx, (J the jacobian, f the residual)
00096         (*pComputeJacobian)(NULL, current_solution, &(linear_system.rGetLhsMatrix()), NULL, NULL, pContext);
00097 
00098         Vec negative_update = linear_system.Solve();
00099 
00100 
00101         Vec test_vec;
00102         VecDuplicate(initialGuess, &test_vec);
00103 
00104         double best_damping_factor = 1.0;
00105         double best_scaled_residual = 1e20; // large
00106 
00107         // Loop over all the possible damping value and determine which gives smallest residual
00108         for (unsigned i=0; i<mTestDampingValues.size(); i++)
00109         {
00110             // Note: WAXPY calls VecWAXPY(w,a,x,y) which computes w = ax+y
00111             PetscVecTools::WAXPY(test_vec,-mTestDampingValues[i],negative_update,current_solution);
00112 
00113             // Compute new residual
00114             linear_system.ZeroLinearSystem();
00115             (*pComputeResidual)(NULL, test_vec, linear_system.rGetRhsVector(), pContext);
00116             VecNorm(linear_system.rGetRhsVector(), NORM_2, &residual_norm);
00117             scaled_residual_norm = residual_norm/size;
00118 
00119             if (scaled_residual_norm < best_scaled_residual)
00120             {
00121                 best_scaled_residual = scaled_residual_norm;
00122                 best_damping_factor = mTestDampingValues[i];
00123             }
00124         }
00125         PetscTools::Destroy(test_vec);
00126 
00127         // Check the smallest residual was actually smaller than the previous; if not, quit
00128         if (best_scaled_residual > old_scaled_residual_norm)
00129         {
00130             // Free memory
00131             PetscTools::Destroy(current_solution);
00132             PetscTools::Destroy(negative_update);
00133 
00134             // Raise error
00135             EXCEPTION("Iteration " << counter << ", unable to find damping factor such that residual decreases in update direction");
00136         }
00137 
00138         if (mWriteStats)
00139         {
00140             std::cout << "    Best damping factor = " << best_damping_factor << "\n";
00141         }
00142 
00143         // Update solution: current_guess = current_solution - best_damping_factor*negative_update
00144         PetscVecTools::AddScaledVector(current_solution, negative_update, -best_damping_factor);
00145         scaled_residual_norm = best_scaled_residual;
00146         PetscTools::Destroy(negative_update);
00147 
00148         // Compute best residual vector again and store in linear_system for next Solve()
00149         linear_system.ZeroLinearSystem();
00150         (*pComputeResidual)(NULL, current_solution, linear_system.rGetRhsVector(), pContext);
00151 
00152         if (mWriteStats)
00153         {
00154             std::cout << "    Iteration " << counter << ": ||residual||/N = " << scaled_residual_norm << "\n";
00155         }
00156     }
00157 
00158     if (mWriteStats)
00159     {
00160         std::cout << "  ..solved!\n\n";
00161     }
00162 
00163     return current_solution;
00164 }
00165 
00166 void SimpleNewtonNonlinearSolver::SetTolerance(double tolerance)
00167 {
00168     assert(tolerance > 0);
00169     mTolerance = tolerance;
00170 }