xref: /petsc/src/tao/bound/impls/bnk/bntr.c (revision 28017e9fc366e9d0379085f0860b391dd2763ed7)
1 #include <../src/tao/bound/impls/bnk/bnk.h>
2 #include <petscksp.h>
3 
4 /*
5  Implements Newton's Method with a trust region approach for solving
6  bound constrained minimization problems.
7 
8  The linear system solve has to be done with a conjugate gradient method.
9 */
10 
11 static PetscErrorCode TaoSolve_BNTR(Tao tao)
12 {
13   PetscErrorCode               ierr;
14   TAO_BNK                      *bnk = (TAO_BNK *)tao->data;
15   KSPConvergedReason           ksp_reason;
16 
17   PetscReal                    oldTrust, prered, actred, stepNorm, steplen;
18   PetscBool                    stepAccepted = PETSC_TRUE;
19   PetscInt                     stepType = BNK_NEWTON;
20 
21   PetscFunctionBegin;
22   /* Initialize the preconditioner, KSP solver and trust radius/line search */
23   tao->reason = TAO_CONTINUE_ITERATING;
24   ierr = TaoBNKInitialize(tao);CHKERRQ(ierr);
25   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
26 
27   /* Have not converged; continue with Newton method */
28   while (tao->reason == TAO_CONTINUE_ITERATING) {
29 
30     if (stepAccepted) {
31       tao->niter++;
32       tao->ksp_its=0;
33       /* Compute the Hessian */
34       ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr);
35       /* Update the BFGS preconditioner */
36       if (BNK_PC_BFGS == bnk->pc_type) {
37         if (BFGS_SCALE_PHESS == bnk->bfgs_scale_type) {
38           /* Obtain diagonal for the bfgs preconditioner  */
39           ierr = MatGetDiagonal(tao->hessian, bnk->Diag);CHKERRQ(ierr);
40           ierr = VecAbs(bnk->Diag);CHKERRQ(ierr);
41           ierr = VecReciprocal(bnk->Diag);CHKERRQ(ierr);
42           ierr = MatLMVMSetScale(bnk->M,bnk->Diag);CHKERRQ(ierr);
43         }
44         /* Update the limited memory preconditioner and get existing # of updates */
45         ierr = MatLMVMUpdate(bnk->M, tao->solution, bnk->unprojected_gradient);CHKERRQ(ierr);
46       }
47     }
48 
49     /* Use the common BNK kernel to compute the Newton step (for inactive variables only) */
50     ierr = TaoBNKComputeStep(tao, &ksp_reason);CHKERRQ(ierr);
51 
52     /* Store current solution before it changes */
53     oldTrust = tao->trust;
54     bnk->fold = bnk->f;
55     ierr = VecCopy(tao->solution, bnk->Xold);CHKERRQ(ierr);
56     ierr = VecCopy(tao->gradient, bnk->Gold);CHKERRQ(ierr);
57     ierr = VecCopy(bnk->unprojected_gradient, bnk->unprojected_gradient_old);CHKERRQ(ierr);
58 
59     /* Temporarily accept the step and project it into the bounds */
60     ierr = VecAXPY(tao->solution, 1.0, tao->stepdirection);CHKERRQ(ierr);
61     ierr = VecMedian(tao->XL, tao->solution, tao->XU, tao->solution);CHKERRQ(ierr);
62 
63     /* Check if the projection changed the step direction */
64     ierr = VecCopy(tao->solution, tao->stepdirection);CHKERRQ(ierr);
65     ierr = VecAXPY(tao->stepdirection, -1.0, bnk->Xold);CHKERRQ(ierr);
66     ierr = VecNorm(tao->stepdirection, NORM_2, &stepNorm);CHKERRQ(ierr);
67     if (stepNorm != bnk->dnorm) {
68       /* Projection changed the step, so we have to recompute predicted reduction.
69          However, we deliberately do not change the step norm and the trust radius
70          in order for the safeguard to more closely mimic a piece-wise linesearch
71          along the bounds. */
72       ierr = MatMult(bnk->H_inactive, tao->stepdirection, bnk->Xwork);CHKERRQ(ierr);
73       ierr = VecAYPX(bnk->Xwork, -0.5, tao->gradient);CHKERRQ(ierr);
74       ierr = VecDot(bnk->Xwork, tao->stepdirection, &prered);
75     } else {
76       /* Step did not change, so we can just recover the pre-computed prediction */
77       ierr = KSPCGGetObjFcn(tao->ksp, &prered);CHKERRQ(ierr);
78     }
79     prered = -prered;
80 
81     /* Compute the actual reduction and update the trust radius */
82     ierr = TaoComputeObjective(tao, tao->solution, &bnk->f);CHKERRQ(ierr);
83     actred = bnk->fold - bnk->f;
84     ierr = TaoBNKUpdateTrustRadius(tao, prered, actred, bnk->update_type, stepType, &stepAccepted);CHKERRQ(ierr);
85 
86     if (stepAccepted) {
87       /* Step is good, evaluate the gradient and the hessian */
88       steplen = 1.0;
89       ++bnk->newt;
90       ierr = TaoComputeGradient(tao, tao->solution, bnk->unprojected_gradient);CHKERRQ(ierr);
91       ierr = VecBoundGradientProjection(bnk->unprojected_gradient,tao->solution,tao->XL,tao->XU,tao->gradient);CHKERRQ(ierr);
92     } else {
93       /* Step is bad, revert old solution and re-solve with new radius*/
94       steplen = 0.0;
95       bnk->f = bnk->fold;
96       ierr = VecCopy(bnk->Xold, tao->solution);CHKERRQ(ierr);
97       ierr = VecCopy(bnk->Gold, tao->gradient);CHKERRQ(ierr);
98       ierr = VecCopy(bnk->unprojected_gradient_old, bnk->unprojected_gradient);CHKERRQ(ierr);
99       if (oldTrust == tao->trust) {
100         /* Can't change the radius anymore so just terminate */
101         tao->reason = TAO_DIVERGED_TR_REDUCTION;
102       }
103     }
104 
105     /*  Check for termination */
106     ierr = TaoGradientNorm(tao, tao->gradient, NORM_2, &bnk->gnorm);CHKERRQ(ierr);
107     if (PetscIsInfOrNanReal(bnk->gnorm)) SETERRQ(PETSC_COMM_SELF,1,"User provided compute function generated Not-a-Number");
108     ierr = TaoLogConvergenceHistory(tao,bnk->f,bnk->gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
109     ierr = TaoMonitor(tao,tao->niter,bnk->f,bnk->gnorm,0.0,steplen);CHKERRQ(ierr);
110     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
111   }
112   PetscFunctionReturn(0);
113 }
114 
115 /*------------------------------------------------------------*/
116 
117 PETSC_EXTERN PetscErrorCode TaoCreate_BNTR(Tao tao)
118 {
119   TAO_BNK        *bnk;
120   PetscErrorCode ierr;
121 
122   PetscFunctionBegin;
123   ierr = TaoCreate_BNK(tao);CHKERRQ(ierr);
124   tao->ops->solve=TaoSolve_BNTR;
125 
126   bnk = (TAO_BNK *)tao->data;
127   bnk->update_type = BNK_UPDATE_REDUCTION; /* trust region updates based on predicted/actual reduction */
128   bnk->sval = 0.0; /* disable Hessian shifting */
129   PetscFunctionReturn(0);
130 }