xref: /petsc/src/tao/bound/impls/bnk/bnls.c (revision bcd4bb4a4158aa96f212e9537e87b40407faf83e)
1 #include <../src/tao/bound/impls/bnk/bnk.h>
2 #include <petscksp.h>
3 
4 /*
5  Implements Newton's Method with a line search approach for
6  solving bound constrained minimization problems.
7 
8  ------------------------------------------------------------
9 
10  x_0 = VecMedian(x_0)
11  f_0, g_0 = TaoComputeObjectiveAndGradient(x_0)
12  pg_0 = project(g_0)
13  check convergence at pg_0
14  needH = TaoBNKInitialize(default:BNK_INIT_DIRECTION)
15  niter = 0
16  step_accepted = true
17 
18  while niter < max_it
19     if needH
20       If max_cg_steps > 0
21         x_k, g_k, pg_k = TaoSolve(BNCG)
22       end
23 
24       H_k = TaoComputeHessian(x_k)
25       if pc_type == BNK_PC_BFGS
26         add correction to BFGS approx
27         if scale_type == BNK_SCALE_AHESS
28           D = VecMedian(1e-6, abs(diag(H_k)), 1e6)
29           scale BFGS with VecReciprocal(D)
30         end
31       end
32       needH = False
33     end
34 
35     if pc_type = BNK_PC_BFGS
36       B_k = BFGS
37     else
38       B_k = VecMedian(1e-6, abs(diag(H_k)), 1e6)
39       B_k = VecReciprocal(B_k)
40     end
41     w = x_k - VecMedian(x_k - 0.001*B_k*g_k)
42     eps = min(eps, norm2(w))
43     determine the active and inactive index sets such that
44       L = {i : (x_k)_i <= l_i + eps && (g_k)_i > 0}
45       U = {i : (x_k)_i >= u_i - eps && (g_k)_i < 0}
46       F = {i : l_i = (x_k)_i = u_i}
47       A = {L + U + F}
48       IA = {i : i not in A}
49 
50     generate the reduced system Hr_k dr_k = -gr_k for variables in IA
51     if p > 0
52       Hr_k += p*
53     end
54     if pc_type == BNK_PC_BFGS && scale_type == BNK_SCALE_PHESS
55       D = VecMedian(1e-6, abs(diag(Hr_k)), 1e6)
56       scale BFGS with VecReciprocal(D)
57     end
58     solve Hr_k dr_k = -gr_k
59     set d_k to (l - x) for variables in L, (u - x) for variables in U, and 0 for variables in F
60 
61     if dot(d_k, pg_k)) >= 0 || norm(d_k) == NaN || norm(d_k) == Inf
62       dr_k = -BFGS*gr_k for variables in I
63       if dot(d_k, pg_k)) >= 0 || norm(d_k) == NaN || norm(d_k) == Inf
64         reset the BFGS preconditioner
65         calculate scale delta and apply it to BFGS
66         dr_k = -BFGS*gr_k for variables in I
67         if dot(d_k, pg_k)) >= 0 || norm(d_k) == NaN || norm(d_k) == Inf
68           dr_k = -gr_k for variables in I
69         end
70       end
71     end
72 
73     x_{k+1}, f_{k+1}, g_{k+1}, ls_failed = TaoBNKPerformLineSearch()
74     if ls_failed
75       f_{k+1} = f_k
76       x_{k+1} = x_k
77       g_{k+1} = g_k
78       pg_{k+1} = pg_k
79       terminate
80     else
81       pg_{k+1} = project(g_{k+1})
82       count the accepted step type (Newton, BFGS, scaled grad or grad)
83     end
84 
85     niter += 1
86     check convergence at pg_{k+1}
87  end
88 */
89 
90 PetscErrorCode TaoSolve_BNLS(Tao tao)
91 {
92   TAO_BNK                     *bnk = (TAO_BNK *)tao->data;
93   KSPConvergedReason           ksp_reason;
94   TaoLineSearchConvergedReason ls_reason;
95   PetscReal                    steplen = 1.0, resnorm;
96   PetscBool                    cgTerminate, needH = PETSC_TRUE, stepAccepted, shift = PETSC_TRUE;
97   PetscInt                     stepType;
98 
99   PetscFunctionBegin;
100   /* Initialize the preconditioner, KSP solver and trust radius/line search */
101   tao->reason = TAO_CONTINUE_ITERATING;
102   PetscCall(TaoBNKInitialize(tao, bnk->init_type, &needH));
103   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(PETSC_SUCCESS);
104 
105   /* Have not converged; continue with Newton method */
106   while (tao->reason == TAO_CONTINUE_ITERATING) {
107     /* Call general purpose update function */
108     if (tao->ops->update) {
109       PetscUseTypeMethod(tao, update, tao->niter, tao->user_update);
110       PetscCall(TaoComputeObjective(tao, tao->solution, &bnk->f));
111     }
112 
113     if (needH && bnk->inactive_idx) {
114       /* Take BNCG steps (if enabled) to trade-off Hessian evaluations for more gradient evaluations */
115       PetscCall(TaoBNKTakeCGSteps(tao, &cgTerminate));
116       if (cgTerminate) {
117         tao->reason = bnk->bncg->reason;
118         PetscFunctionReturn(PETSC_SUCCESS);
119       }
120       /* Compute the hessian and update the BFGS preconditioner at the new iterate */
121       PetscCall((*bnk->computehessian)(tao));
122       needH = PETSC_FALSE;
123     }
124 
125     /* Use the common BNK kernel to compute the safeguarded Newton step (for inactive variables only) */
126     PetscCall((*bnk->computestep)(tao, shift, &ksp_reason, &stepType));
127     PetscCall(TaoBNKSafeguardStep(tao, ksp_reason, &stepType));
128 
129     /* Store current solution before it changes */
130     bnk->fold = bnk->f;
131     PetscCall(VecCopy(tao->solution, bnk->Xold));
132     PetscCall(VecCopy(tao->gradient, bnk->Gold));
133     PetscCall(VecCopy(bnk->unprojected_gradient, bnk->unprojected_gradient_old));
134 
135     /* Trigger the line search */
136     PetscCall(TaoBNKPerformLineSearch(tao, &stepType, &steplen, &ls_reason));
137 
138     if (ls_reason != TAOLINESEARCH_SUCCESS && ls_reason != TAOLINESEARCH_SUCCESS_USER) {
139       /* Failed to find an improving point */
140       needH  = PETSC_FALSE;
141       bnk->f = bnk->fold;
142       PetscCall(VecCopy(bnk->Xold, tao->solution));
143       PetscCall(VecCopy(bnk->Gold, tao->gradient));
144       PetscCall(VecCopy(bnk->unprojected_gradient_old, bnk->unprojected_gradient));
145       steplen     = 0.0;
146       tao->reason = TAO_DIVERGED_LS_FAILURE;
147     } else {
148       /* new iterate so we need to recompute the Hessian */
149       needH = PETSC_TRUE;
150       /* compute the projected gradient */
151       PetscCall(TaoBNKEstimateActiveSet(tao, bnk->as_type));
152       PetscCall(VecCopy(bnk->unprojected_gradient, tao->gradient));
153       if (bnk->active_idx) PetscCall(VecISSet(tao->gradient, bnk->active_idx, 0.0));
154       PetscCall(TaoGradientNorm(tao, tao->gradient, NORM_2, &bnk->gnorm));
155       /* update the trust radius based on the step length */
156       PetscCall(TaoBNKUpdateTrustRadius(tao, 0.0, 0.0, BNK_UPDATE_STEP, stepType, &stepAccepted));
157       /* count the accepted step type */
158       PetscCall(TaoBNKAddStepCounts(tao, stepType));
159       /* active BNCG recycling for next iteration */
160       PetscCall(TaoSetRecycleHistory(bnk->bncg, PETSC_TRUE));
161     }
162 
163     /*  Check for termination */
164     PetscCall(VecFischer(tao->solution, bnk->unprojected_gradient, tao->XL, tao->XU, bnk->W));
165     PetscCall(VecNorm(bnk->W, NORM_2, &resnorm));
166     PetscCheck(!PetscIsInfOrNanReal(resnorm), PetscObjectComm((PetscObject)tao), PETSC_ERR_USER, "User provided compute function generated infinity or NaN");
167     ++tao->niter;
168     PetscCall(TaoLogConvergenceHistory(tao, bnk->f, resnorm, 0.0, tao->ksp_its));
169     PetscCall(TaoMonitor(tao, tao->niter, bnk->f, resnorm, 0.0, steplen));
170     PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
171   }
172   PetscFunctionReturn(PETSC_SUCCESS);
173 }
174 
175 /*MC
176   TAOBNLS - Bounded Newton Line Search for nonlinear minimization with bound constraints.
177 
178   Options Database Keys:
179 + -tao_bnk_max_cg_its - maximum number of bounded conjugate-gradient iterations taken in each Newton loop
180 . -tao_bnk_init_type - trust radius initialization method ("constant", "direction", "interpolation")
181 . -tao_bnk_update_type - trust radius update method ("step", "direction", "interpolation")
182 - -tao_bnk_as_type - active-set estimation method ("none", "bertsekas")
183 
184   Level: beginner
185 M*/
186 PETSC_EXTERN PetscErrorCode TaoCreate_BNLS(Tao tao)
187 {
188   TAO_BNK *bnk;
189 
190   PetscFunctionBegin;
191   PetscCall(TaoCreate_BNK(tao));
192   tao->ops->solve = TaoSolve_BNLS;
193 
194   bnk              = (TAO_BNK *)tao->data;
195   bnk->init_type   = BNK_INIT_DIRECTION;
196   bnk->update_type = BNK_UPDATE_STEP; /* trust region updates based on line search step length */
197   PetscFunctionReturn(PETSC_SUCCESS);
198 }
199