xref: /petsc/src/tao/bound/impls/bqnls/bqnls.c (revision 6d8694c4fbab79f9439f1ad13c0386ba7ee1ca4b)
1 #include <../src/tao/bound/impls/bqnk/bqnk.h>
2 
3 static const char *BNK_AS[64] = {"none", "bertsekas"};
4 
TaoBQNLSComputeHessian(Tao tao)5 static PetscErrorCode TaoBQNLSComputeHessian(Tao tao)
6 {
7   TAO_BNK  *bnk  = (TAO_BNK *)tao->data;
8   TAO_BQNK *bqnk = (TAO_BQNK *)bnk->ctx;
9   PetscReal gnorm2, delta;
10 
11   PetscFunctionBegin;
12   /* Compute the initial scaling and update the approximation */
13   gnorm2 = bnk->gnorm * bnk->gnorm;
14   if (gnorm2 == 0.0) gnorm2 = PETSC_MACHINE_EPSILON;
15   if (bnk->f == 0.0) delta = 2.0 / gnorm2;
16   else delta = 2.0 * PetscAbsScalar(bnk->f) / gnorm2;
17   PetscCall(MatLMVMSymBroydenSetDelta(bqnk->B, delta));
18   PetscCall(MatLMVMUpdate(bqnk->B, tao->solution, bnk->unprojected_gradient));
19   PetscFunctionReturn(PETSC_SUCCESS);
20 }
21 
TaoBQNLSComputeStep(Tao tao,PetscBool shift,KSPConvergedReason * ksp_reason,PetscInt * step_type)22 static PetscErrorCode TaoBQNLSComputeStep(Tao tao, PetscBool shift, KSPConvergedReason *ksp_reason, PetscInt *step_type)
23 {
24   TAO_BNK  *bnk  = (TAO_BNK *)tao->data;
25   TAO_BQNK *bqnk = (TAO_BQNK *)bnk->ctx;
26   PetscInt  nupdates;
27 
28   PetscFunctionBegin;
29   PetscCall(MatSolve(bqnk->B, tao->gradient, tao->stepdirection));
30   PetscCall(VecScale(tao->stepdirection, -1.0));
31   PetscCall(TaoBNKBoundStep(tao, bnk->as_type, tao->stepdirection));
32   *ksp_reason = KSP_CONVERGED_ATOL;
33   PetscCall(MatLMVMGetUpdateCount(bqnk->B, &nupdates));
34   if (nupdates == 0) *step_type = BNK_SCALED_GRADIENT;
35   else *step_type = BNK_BFGS;
36   PetscFunctionReturn(PETSC_SUCCESS);
37 }
38 
TaoSetFromOptions_BQNLS(Tao tao,PetscOptionItems PetscOptionsObject)39 static PetscErrorCode TaoSetFromOptions_BQNLS(Tao tao, PetscOptionItems PetscOptionsObject)
40 {
41   TAO_BNK  *bnk  = (TAO_BNK *)tao->data;
42   TAO_BQNK *bqnk = (TAO_BQNK *)bnk->ctx;
43   PetscBool is_set, is_spd;
44 
45   PetscFunctionBegin;
46   PetscOptionsHeadBegin(PetscOptionsObject, "Quasi-Newton-Krylov method for bound constrained optimization");
47   PetscCall(PetscOptionsEList("-tao_bnk_as_type", "active set estimation method", "", BNK_AS, BNK_AS_TYPES, BNK_AS[bnk->as_type], &bnk->as_type, NULL));
48   PetscCall(PetscOptionsReal("-tao_bnk_epsilon", "(developer) tolerance used when computing actual and predicted reduction", "", bnk->epsilon, &bnk->epsilon, NULL));
49   PetscCall(PetscOptionsReal("-tao_bnk_as_tol", "(developer) initial tolerance used when estimating actively bounded variables", "", bnk->as_tol, &bnk->as_tol, NULL));
50   PetscCall(PetscOptionsReal("-tao_bnk_as_step", "(developer) step length used when estimating actively bounded variables", "", bnk->as_step, &bnk->as_step, NULL));
51   PetscCall(PetscOptionsInt("-tao_bnk_max_cg_its", "number of BNCG iterations to take for each Newton step", "", bnk->max_cg_its, &bnk->max_cg_its, NULL));
52   PetscOptionsHeadEnd();
53 
54   PetscCall(TaoSetOptionsPrefix(bnk->bncg, ((PetscObject)tao)->prefix));
55   PetscCall(TaoAppendOptionsPrefix(bnk->bncg, "tao_bnk_"));
56   PetscCall(TaoSetFromOptions(bnk->bncg));
57 
58   PetscCall(MatSetOptionsPrefix(bqnk->B, ((PetscObject)tao)->prefix));
59   PetscCall(MatAppendOptionsPrefix(bqnk->B, "tao_bqnls_"));
60   PetscCall(MatSetFromOptions(bqnk->B));
61   PetscCall(MatIsSPDKnown(bqnk->B, &is_set, &is_spd));
62   PetscCheck(is_set && is_spd, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_INCOMP, "LMVM matrix must be symmetric positive-definite");
63   PetscFunctionReturn(PETSC_SUCCESS);
64 }
65 
66 /*MC
67   TAOBQNLS - Bounded Quasi-Newton Line Search method for nonlinear minimization with bound
68              constraints. This method approximates the action of the inverse-Hessian with a
69              limited memory quasi-Newton formula. The quasi-Newton matrix and its options are
70              accessible via the prefix `-tao_bqnls_`
71 
72   Options Database Keys:
73 + -tao_bnk_max_cg_its - maximum number of bounded conjugate-gradient iterations taken in each Newton loop
74 . -tao_bnk_as_type - active-set estimation method ("none", "bertsekas")
75 . -tao_bnk_epsilon - (developer) tolerance for small pred/actual ratios that trigger automatic step acceptance
76 . -tao_bnk_as_tol - (developer) initial tolerance used in estimating bounded active variables (-as_type bertsekas)
77 - -tao_bnk_as_step - (developer) trial step length used in estimating bounded active variables (-as_type bertsekas)
78 
79   Level: beginner
80 
81 .seealso: `TAOBNK`
82 M*/
TaoCreate_BQNLS(Tao tao)83 PETSC_EXTERN PetscErrorCode TaoCreate_BQNLS(Tao tao)
84 {
85   TAO_BNK  *bnk;
86   TAO_BQNK *bqnk;
87 
88   PetscFunctionBegin;
89   PetscCall(TaoCreate_BQNK(tao));
90   tao->ops->setfromoptions = TaoSetFromOptions_BQNLS;
91 
92   bnk                 = (TAO_BNK *)tao->data;
93   bnk->update_type    = BNK_UPDATE_STEP;
94   bnk->computehessian = TaoBQNLSComputeHessian;
95   bnk->computestep    = TaoBQNLSComputeStep;
96 
97   bqnk        = (TAO_BQNK *)bnk->ctx;
98   bqnk->solve = TaoSolve_BNLS;
99   PetscCall(MatSetType(bqnk->B, MATLMVMBFGS));
100   PetscFunctionReturn(PETSC_SUCCESS);
101 }
102