xref: /petsc/src/tao/bound/impls/bnk/bnk.h (revision 62675bee48cd89cd8757a6de6d7b87f2edd3afcc)
1 /*
2 Context for bounded Newton-Krylov type optimization algorithms
3 */
4 
5 #if !defined(__TAO_BNK_H)
6 #define __TAO_BNK_H
7 #include <petsc/private/taoimpl.h>
8 #include <../src/tao/matrix/lmvmmat.h>
9 
10 typedef struct {
11   Mat H_inactive, Hpre_inactive, M;
12   Vec W, Xwork, Gwork, inactive_work, active_work;
13   Vec Xold, Gold, Diag, Diag_min, Diag_max;
14   Vec unprojected_gradient, unprojected_gradient_old;
15   IS  inactive_idx, active_idx, active_lower, active_upper, active_fixed;
16 
17   /* Scalar values for the solution and step */
18   PetscReal fold, f, gnorm, dnorm;
19 
20   /* Parameters for active set estimation */
21   PetscReal as_tol;
22   PetscReal as_step;
23 
24   /* Parameters when updating the perturbation added to the Hessian matrix
25      according to the following scheme:
26 
27      pert = sval;
28 
29      do until convergence
30        shift Hessian by pert
31        solve Newton system
32 
33        if (linear solver failed or did not compute a descent direction)
34          use steepest descent direction and increase perturbation
35 
36          if (0 == pert)
37            initialize perturbation
38            pert = min(imax, max(imin, imfac * norm(G)))
39          else
40            increase perturbation
41            pert = min(pmax, max(pgfac * pert, pmgfac * norm(G)))
42          fi
43        else
44          use linear solver direction and decrease perturbation
45 
46          pert = min(psfac * pert, pmsfac * norm(G))
47          if (pert < pmin)
48            pert = 0
49          fi
50        fi
51 
52        perform line search
53        function and gradient evaluation
54        check convergence
55      od
56   */
57   PetscReal sval;               /*  Starting perturbation value, default zero */
58 
59   PetscReal imin;               /*  Minimum perturbation added during initialization  */
60   PetscReal imax;               /*  Maximum perturbation added during initialization */
61   PetscReal imfac;              /*  Merit function factor during initialization */
62 
63   PetscReal pert;               /*  Current perturbation value */
64   PetscReal pmin;               /*  Minimim perturbation value */
65   PetscReal pmax;               /*  Maximum perturbation value */
66   PetscReal pgfac;              /*  Perturbation growth factor */
67   PetscReal psfac;              /*  Perturbation shrink factor */
68   PetscReal pmgfac;             /*  Merit function growth factor */
69   PetscReal pmsfac;             /*  Merit function shrink factor */
70 
71   /* Parameters when updating the trust-region radius based on steplength
72      if   step < nu1            (very bad step)
73        radius = omega1 * min(norm(d), radius)
74      elif step < nu2            (bad step)
75        radius = omega2 * min(norm(d), radius)
76      elif step < nu3            (okay step)
77        radius = omega3 * radius;
78      elif step < nu4            (good step)
79        radius = max(omega4 * norm(d), radius)
80      else                       (very good step)
81        radius = max(omega5 * norm(d), radius)
82      fi
83   */
84   PetscReal nu1;                /*  used to compute trust-region radius */
85   PetscReal nu2;                /*  used to compute trust-region radius */
86   PetscReal nu3;                /*  used to compute trust-region radius */
87   PetscReal nu4;                /*  used to compute trust-region radius */
88 
89   PetscReal omega1;             /*  factor used for trust-region update */
90   PetscReal omega2;             /*  factor used for trust-region update */
91   PetscReal omega3;             /*  factor used for trust-region update */
92   PetscReal omega4;             /*  factor used for trust-region update */
93   PetscReal omega5;             /*  factor used for trust-region update */
94 
95   /* Parameters when updating the trust-region radius based on reduction
96 
97      kappa = ared / pred
98      if   kappa < eta1          (very bad step)
99        radius = alpha1 * min(norm(d), radius)
100      elif kappa < eta2          (bad step)
101        radius = alpha2 * min(norm(d), radius)
102      elif kappa < eta3          (okay step)
103        radius = alpha3 * radius;
104      elif kappa < eta4          (good step)
105        radius = max(alpha4 * norm(d), radius)
106      else                       (very good step)
107        radius = max(alpha5 * norm(d), radius)
108      fi
109   */
110   PetscReal eta1;               /*  used to compute trust-region radius */
111   PetscReal eta2;               /*  used to compute trust-region radius */
112   PetscReal eta3;               /*  used to compute trust-region radius */
113   PetscReal eta4;               /*  used to compute trust-region radius */
114 
115   PetscReal alpha1;             /*  factor used for trust-region update */
116   PetscReal alpha2;             /*  factor used for trust-region update */
117   PetscReal alpha3;             /*  factor used for trust-region update */
118   PetscReal alpha4;             /*  factor used for trust-region update */
119   PetscReal alpha5;             /*  factor used for trust-region update */
120 
121   /* Parameters when updating the trust-region radius based on interpolation
122 
123      kappa = ared / pred
124      if   kappa >= 1.0 - mu1    (very good step)
125        choose tau in [gamma3, gamma4]
126        radius = max(tau * norm(d), radius)
127      elif kappa >= 1.0 - mu2    (good step)
128        choose tau in [gamma2, gamma3]
129        if (tau >= 1.0)
130          radius = max(tau * norm(d), radius)
131        else
132          radius = tau * min(norm(d), radius)
133        fi
134      else                       (bad step)
135        choose tau in [gamma1, 1.0]
136        radius = tau * min(norm(d), radius)
137      fi
138   */
139   PetscReal mu1;                /*  used for model agreement in interpolation */
140   PetscReal mu2;                /*  used for model agreement in interpolation */
141 
142   PetscReal gamma1;             /*  factor used for interpolation */
143   PetscReal gamma2;             /*  factor used for interpolation */
144   PetscReal gamma3;             /*  factor used for interpolation */
145   PetscReal gamma4;             /*  factor used for interpolation */
146 
147   PetscReal theta;              /*  factor used for interpolation */
148 
149   /*  Parameters when initializing trust-region radius based on interpolation */
150   PetscReal mu1_i;              /*  used for model agreement in interpolation */
151   PetscReal mu2_i;              /*  used for model agreement in interpolation */
152 
153   PetscReal gamma1_i;           /*  factor used for interpolation */
154   PetscReal gamma2_i;           /*  factor used for interpolation */
155   PetscReal gamma3_i;           /*  factor used for interpolation */
156   PetscReal gamma4_i;           /*  factor used for interpolation */
157 
158   PetscReal theta_i;            /*  factor used for interpolation */
159 
160   /*  Other parameters */
161   PetscReal min_radius;         /*  lower bound on initial radius value */
162   PetscReal max_radius;         /*  upper bound on trust region radius */
163   PetscReal epsilon;            /*  tolerance used when computing ared/pred */
164   PetscReal dmin, dmax;         /*  upper and lower bounds for the Hessian diagonal vector */
165 
166   PetscInt newt;                /*  Newton directions attempted */
167   PetscInt bfgs;                /*  BFGS directions attempted */
168   PetscInt sgrad;               /*  Scaled gradient directions attempted */
169   PetscInt grad;                /*  Gradient directions attempted */
170 
171   PetscInt as_type;             /*   Active set estimation method */
172   PetscInt pc_type;             /*  Preconditioner for the code */
173   PetscInt bfgs_scale_type;     /*  Scaling matrix to used for the bfgs preconditioner */
174   PetscInt init_type;           /*  Trust-region initialization method */
175   PetscInt update_type;         /*  Trust-region update method */
176 
177   /* Trackers for KSP solution type and convergence reasons */
178   PetscInt ksp_atol;
179   PetscInt ksp_rtol;
180   PetscInt ksp_ctol;
181   PetscInt ksp_negc;
182   PetscInt ksp_dtol;
183   PetscInt ksp_iter;
184   PetscInt ksp_othr;
185   PetscBool is_nash, is_stcg, is_gltr;
186 } TAO_BNK;
187 
188 #endif /* if !defined(__TAO_BNK_H) */
189 
190 #define BNK_NEWTON              0
191 #define BNK_BFGS                1
192 #define BNK_SCALED_GRADIENT     2
193 #define BNK_GRADIENT            3
194 
195 #define BNK_PC_NONE     0
196 #define BNK_PC_AHESS    1
197 #define BNK_PC_BFGS     2
198 #define BNK_PC_PETSC    3
199 #define BNK_PC_TYPES    4
200 
201 #define BFGS_SCALE_AHESS        0
202 #define BFGS_SCALE_PHESS        1
203 #define BFGS_SCALE_BFGS         2
204 #define BFGS_SCALE_TYPES        3
205 
206 #define BNK_INIT_CONSTANT         0
207 #define BNK_INIT_DIRECTION        1
208 #define BNK_INIT_INTERPOLATION    2
209 #define BNK_INIT_TYPES            3
210 
211 #define BNK_UPDATE_STEP           0
212 #define BNK_UPDATE_REDUCTION      1
213 #define BNK_UPDATE_INTERPOLATION  2
214 #define BNK_UPDATE_TYPES          3
215 
216 #define BNK_AS_NONE             0
217 #define BNK_AS_BERTSEKAS        1
218 #define BNK_AS_TYPES            2
219 
220 static const char *BNK_PC[64] = {"none", "ahess", "bfgs", "petsc"};
221 
222 static const char *BFGS_SCALE[64] = {"ahess", "phess", "bfgs"};
223 
224 static const char *BNK_INIT[64] = {"constant", "direction", "interpolation"};
225 
226 static const char *BNK_UPDATE[64] = {"step", "reduction", "interpolation"};
227 
228 static const char *BNK_AS[64] = {"none", "bertsekas"};
229 
230 PETSC_INTERN PetscErrorCode TaoCreate_BNK(Tao);
231 
232 PETSC_INTERN PetscErrorCode MatLMVMSolveShell(PC, Vec, Vec);
233 PETSC_INTERN PetscErrorCode TaoBNKInitialize(Tao, PetscInt);
234 PETSC_INTERN PetscErrorCode TaoBNKEstimateActiveSet(Tao);
235 PETSC_INTERN PetscErrorCode TaoBNKComputeHessian(Tao);
236 PETSC_INTERN PetscErrorCode TaoBNKBoundStep(Tao, Vec);
237 PETSC_INTERN PetscErrorCode TaoBNKComputeStep(Tao, PetscBool, KSPConvergedReason*);
238 PETSC_INTERN PetscErrorCode TaoBNKSafeguardStep(Tao, KSPConvergedReason, PetscInt*);
239 PETSC_INTERN PetscErrorCode TaoBNKPerformLineSearch(Tao, PetscInt, PetscReal*, TaoLineSearchConvergedReason*);
240 PETSC_INTERN PetscErrorCode TaoBNKUpdateTrustRadius(Tao, PetscReal, PetscReal, PetscInt, PetscInt, PetscBool*);
241 PETSC_INTERN PetscErrorCode TaoBNKAddStepCounts(Tao, PetscInt);
242