00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026
00027
00028
00029
00030
00031
00032
00033
00034
00035
00036
00037
00038
00039
00040
00041
00042
00043
00044
00045
00046
00047
00048
00049
00050 #ifndef LBFGS_H_
00051 #define LBFGS_H_
00052
00053 #include "fastlib/fastlib.h"
00054 #include <string>
00055
00070 const fx_entry_doc lbfgs_entries[] = {
00071 {"num_of_points", FX_PARAM, FX_INT, NULL,
00072 " The number of points for the optimization variable.\n"},
00073 {"sigma", FX_PARAM, FX_DOUBLE, NULL,
00074 " The initial penalty parameter on the augmented lagrangian.\n"},
00075 {"objective_factor", FX_PARAM, FX_DOUBLE, NULL,
00076 " obsolete.\n"},
00077 {"eta", FX_PARAM, FX_DOUBLE, NULL,
00078 " wolfe parameter.\n"},
00079 {"gamma", FX_PARAM, FX_DOUBLE, NULL,
00080 " sigma increase rate, after inner loop is done sigma is multiplied by gamma.\n"},
00081 {"new_dimension", FX_PARAM, FX_INT, NULL,
00082 " The dimension of the points\n"},
00083 {"desired_feasibility", FX_PARAM, FX_DOUBLE, NULL,
00084 " Since this is used with augmented lagrangian, we need to know "
00085 "when the feasibility is sufficient.\n"},
00086 {"feasibility_tolerance", FX_PARAM, FX_DOUBLE, NULL,
00087 " if the feasibility is not improved by that quantity, then it stops.\n"},
00088 {"wolfe_sigma1", FX_PARAM, FX_DOUBLE, NULL,
00089 " wolfe parameter.\n"},
00090 {"wolfe_sigma2", FX_PARAM, FX_DOUBLE, NULL,
00091 " wolfe parameter.\n"},
00092 {"min_beta", FX_PARAM, FX_DOUBLE, NULL,
00093 " wolfe parameter.\n"},
00094 {"wolfe_beta", FX_PARAM, FX_DOUBLE, NULL,
00095 " wolfe parameter.\n"},
00096 {"step_size", FX_PARAM, FX_DOUBLE, NULL,
00097 " Initial step size for the wolfe search.\n"},
00098 {"silent", FX_PARAM, FX_BOOL, NULL,
00099 " if true then it doesn't emmit updates.\n"},
00100 {"show_warnings", FX_PARAM, FX_BOOL, NULL,
00101 " if true then it does show warnings.\n"},
00102 {"use_default_termination", FX_PARAM, FX_BOOL, NULL,
00103 " let this module decide where to terminate. If false then"
00104 " the objective function decides .\n"},
00105 {"norm_grad_tolerance", FX_PARAM, FX_DOUBLE, NULL,
00106 " If the norm of the gradient doesn't change more than "
00107 "this quantity between two iterations and the use_default_termination "
00108 "is set, the algorithm terminates.\n"},
00109 {"max_iterations", FX_PARAM, FX_INT, NULL,
00110 " maximum number of iterations required.\n"},
00111 {"mem_bfgs", FX_PARAM, FX_INT, NULL,
00112 " the limited memory of BFGS.\n"},
00113 {"log_file", FX_PARAM, FX_STR, NULL,
00114 " file to log the output.\n"},
00115 {"iterations", FX_RESULT, FX_INT, NULL,
00116 " iterations until convergence.\n"},
00117 {"feasibility_error", FX_RESULT, FX_DOUBLE, NULL,
00118 " the fesibility error achived by termination.\n"},
00119 {"final_sigma", FX_RESULT, FX_DOUBLE, NULL,
00120 " the last penalty parameter used\n"},
00121 {"objective", FX_RESULT, FX_DOUBLE, NULL,
00122 " the objective achieved by termination.\n"},
00123 {"wolfe_step", FX_TIMER, FX_CUSTOM, NULL,
00124 " Time spent computing the wolfe step.\n"},
00125 {"bfgs_step", FX_TIMER, FX_CUSTOM, NULL,
00126 " Time spent computing the bfgs step.\n"},
00127 {"update_bfgs", FX_TIMER, FX_CUSTOM, NULL,
00128 " Time spent computing the bfgs updating.\n"},
00129
00130 FX_ENTRY_DOC_DONE
00131 };
00132
00133 const fx_module_doc lbfgs_doc = {
00134 lbfgs_entries, NULL,
00135 "The LBFGS module for optimization.\n"
00136 };
00137
00138
00139 template<typename OptimizedFunction>
00140 class Lbfgs {
00141 public:
00142 void Init(OptimizedFunction *optimized_function, datanode* module);
00143 void Destruct();
00144 void ComputeLocalOptimumBFGS();
00145 void ReportProgress();
00146 void ReportProgressFile(std::string file);
00147 void CopyCoordinates(Matrix *result);
00148 void Reset();
00149 void set_coordinates(Matrix &coordinates);
00150 void set_desired_feasibility(double desired_feasibility);
00151 void set_feasibility_tolerance(double feasibility_tolerance);
00152 void set_norm_grad_tolerance(double norm_grad_tolerance);
00153 void set_max_iterations(index_t max_iterations);
00154 Matrix *coordinates();
00155 double sigma();
00156 void set_sigma(double sigma);
00157
00158 private:
00159 void InitOptimization_();
00160 void ComputeWolfeStep_();
00161 void UpdateLagrangeMult_();
00162 success_t ComputeWolfeStep_(double *step, Matrix &direction);
00163 success_t ComputeBFGS_(double *step, Matrix &grad, index_t memory);
00164 success_t UpdateBFGS_();
00165 success_t UpdateBFGS_(index_t index_bfgs);
00166 void BoundConstrain();
00167 std::string ComputeProgress_();
00168 void ReportProgressFile_();
00169
00170 struct datanode* module_;
00171 OptimizedFunction *optimized_function_;
00172 index_t num_of_iterations_;
00173 index_t num_of_points_;
00174 index_t new_dimension_;
00175 double sigma_;
00176 double objective_factor_;
00177 double eta_;
00178 double gamma_;
00179 double step_;
00180 double desired_feasibility_;
00181 double feasibility_tolerance_;
00182 double norm_grad_tolerance_;
00183 double wolfe_sigma1_;
00184 double wolfe_sigma2_;
00185 double wolfe_beta_;
00186 double min_beta_;
00187 bool silent_;
00188 bool show_warnings_;
00189 bool use_default_termination_;
00190 ArrayList<Matrix> s_bfgs_;
00191 ArrayList<Matrix> y_bfgs_;
00192 Vector ro_bfgs_;
00193 index_t index_bfgs_;
00194 Matrix coordinates_;
00195 Matrix previous_coordinates_;
00196 Matrix gradient_;
00197 Matrix previous_gradient_;
00198 index_t max_iterations_;
00199 double step_size_;
00200
00201 index_t mem_bfgs_;
00202 FILE *fp_log_;
00203 };
00204
00205 #include "lbfgs_impl.h"
00206 #endif //LBFGS_H_