1: #include <../src/tao/complementarity/impls/ssls/ssls.h> 2: /*
3: Context for ASXLS
4: -- active-set - reduced matrices formed
5: - inherit properties of original system
6: -- semismooth (S) - function not differentiable
7: - merit function continuously differentiable
8: - Fischer-Burmeister reformulation of complementarity
9: - Billups composition for two finite bounds
10: -- infeasible (I) - iterates not guaranteed to remain within bounds
11: -- feasible (F) - iterates guaranteed to remain within bounds
12: -- linesearch (LS) - Armijo rule on direction
14: Many other reformulations are possible and combinations of
15: feasible/infeasible and linesearch/trust region are possible.
17: Basic theory
18: Fischer-Burmeister reformulation is semismooth with a continuously
19: differentiable merit function and strongly semismooth if the F has
20: lipschitz continuous derivatives.
22: Every accumulation point generated by the algorithm is a stationary
23: point for the merit function. Stationary points of the merit function
24: are solutions of the complementarity problem if
25: a. the stationary point has a BD-regular subdifferential, or
26: b. the Schur complement F'/F'_ff is a P_0-matrix where ff is the
27: index set corresponding to the free variables.
29: If one of the accumulation points has a BD-regular subdifferential then
30: a. the entire sequence converges to this accumulation point at
31: a local q-superlinear rate
32: b. if in addition the reformulation is strongly semismooth near
33: this accumulation point, then the algorithm converges at a
34: local q-quadratic rate.
36: The theory for the feasible version follows from the feasible descent
37: algorithm framework.
39: References:
40: Billups, "Algorithms for Complementarity Problems and Generalized
41: Equations," Ph.D thesis, University of Wisconsin Madison, 1995.
42: De Luca, Facchinei, Kanzow, "A Semismooth Equation Approach to the
43: Solution of Nonlinear Complementarity Problems," Mathematical
44: Programming, 75, pages 407439, 1996.
45: Ferris, Kanzow, Munson, "Feasible Descent Algorithms for Mixed
46: Complementarity Problems," Mathematical Programming, 86,
47: pages 475497, 1999.
48: Fischer, "A Special Newton type Optimization Method," Optimization,
49: 24, 1992
50: Munson, Facchinei, Ferris, Fischer, Kanzow, "The Semismooth Algorithm
51: for Large Scale Complementarity Problems," Technical Report,
52: University of Wisconsin Madison, 1999.
53: */
56: static PetscErrorCode TaoSetUp_ASFLS(Tao tao) 57: {
58: TAO_SSLS *asls = (TAO_SSLS *)tao->data;
62: VecDuplicate(tao->solution,&tao->gradient);
63: VecDuplicate(tao->solution,&tao->stepdirection);
64: VecDuplicate(tao->solution,&asls->ff);
65: VecDuplicate(tao->solution,&asls->dpsi);
66: VecDuplicate(tao->solution,&asls->da);
67: VecDuplicate(tao->solution,&asls->db);
68: VecDuplicate(tao->solution,&asls->t1);
69: VecDuplicate(tao->solution,&asls->t2);
70: VecDuplicate(tao->solution, &asls->w);
71: asls->fixed = NULL;
72: asls->free = NULL;
73: asls->J_sub = NULL;
74: asls->Jpre_sub = NULL;
75: asls->r1 = NULL;
76: asls->r2 = NULL;
77: asls->r3 = NULL;
78: asls->dxfree = NULL;
79: return(0);
80: }
82: static PetscErrorCode Tao_ASLS_FunctionGradient(TaoLineSearch ls, Vec X, PetscReal *fcn, Vec G, void *ptr) 83: {
84: Tao tao = (Tao)ptr;
85: TAO_SSLS *asls = (TAO_SSLS *)tao->data;
89: TaoComputeConstraints(tao, X, tao->constraints);
90: VecFischer(X,tao->constraints,tao->XL,tao->XU,asls->ff);
91: VecNorm(asls->ff,NORM_2,&asls->merit);
92: *fcn = 0.5*asls->merit*asls->merit;
93: TaoComputeJacobian(tao,tao->solution,tao->jacobian,tao->jacobian_pre);
95: MatDFischer(tao->jacobian, tao->solution, tao->constraints,tao->XL, tao->XU, asls->t1, asls->t2,asls->da, asls->db);
96: VecPointwiseMult(asls->t1, asls->ff, asls->db);
97: MatMultTranspose(tao->jacobian,asls->t1,G);
98: VecPointwiseMult(asls->t1, asls->ff, asls->da);
99: VecAXPY(G,1.0,asls->t1);
100: return(0);
101: }
103: static PetscErrorCode TaoDestroy_ASFLS(Tao tao)104: {
105: TAO_SSLS *ssls = (TAO_SSLS *)tao->data;
109: VecDestroy(&ssls->ff);
110: VecDestroy(&ssls->dpsi);
111: VecDestroy(&ssls->da);
112: VecDestroy(&ssls->db);
113: VecDestroy(&ssls->w);
114: VecDestroy(&ssls->t1);
115: VecDestroy(&ssls->t2);
116: VecDestroy(&ssls->r1);
117: VecDestroy(&ssls->r2);
118: VecDestroy(&ssls->r3);
119: VecDestroy(&ssls->dxfree);
120: MatDestroy(&ssls->J_sub);
121: MatDestroy(&ssls->Jpre_sub);
122: ISDestroy(&ssls->fixed);
123: ISDestroy(&ssls->free);
124: PetscFree(tao->data);
125: tao->data = NULL;
126: return(0);
127: }
129: static PetscErrorCode TaoSolve_ASFLS(Tao tao)130: {
131: TAO_SSLS *asls = (TAO_SSLS *)tao->data;
132: PetscReal psi,ndpsi, normd, innerd, t=0;
133: PetscInt nf;
134: PetscErrorCode ierr;
135: TaoLineSearchConvergedReason ls_reason;
138: /* Assume that Setup has been called!
139: Set the structure for the Jacobian and create a linear solver. */
141: TaoComputeVariableBounds(tao);
142: TaoLineSearchSetObjectiveAndGradientRoutine(tao->linesearch,Tao_ASLS_FunctionGradient,tao);
143: TaoLineSearchSetObjectiveRoutine(tao->linesearch,Tao_SSLS_Function,tao);
144: TaoLineSearchSetVariableBounds(tao->linesearch,tao->XL,tao->XU);
146: VecMedian(tao->XL, tao->solution, tao->XU, tao->solution);
148: /* Calculate the function value and fischer function value at the
149: current iterate */
150: TaoLineSearchComputeObjectiveAndGradient(tao->linesearch,tao->solution,&psi,asls->dpsi);
151: VecNorm(asls->dpsi,NORM_2,&ndpsi);
153: tao->reason = TAO_CONTINUE_ITERATING;
154: while (1) {
155: /* Check the converged criteria */
156: PetscInfo3(tao,"iter %D, merit: %g, ||dpsi||: %g\n",tao->niter,(double)asls->merit,(double)ndpsi);
157: TaoLogConvergenceHistory(tao,asls->merit,ndpsi,0.0,tao->ksp_its);
158: TaoMonitor(tao,tao->niter,asls->merit,ndpsi,0.0,t);
159: (*tao->ops->convergencetest)(tao,tao->cnvP);
160: if (TAO_CONTINUE_ITERATING != tao->reason) break;
161: tao->niter++;
163: /* We are going to solve a linear system of equations. We need to
164: set the tolerances for the solve so that we maintain an asymptotic
165: rate of convergence that is superlinear.
166: Note: these tolerances are for the reduced system. We really need
167: to make sure that the full system satisfies the full-space conditions.
169: This rule gives superlinear asymptotic convergence
170: asls->atol = min(0.5, asls->merit*sqrt(asls->merit));
171: asls->rtol = 0.0;
173: This rule gives quadratic asymptotic convergence
174: asls->atol = min(0.5, asls->merit*asls->merit);
175: asls->rtol = 0.0;
177: Calculate a free and fixed set of variables. The fixed set of
178: variables are those for the d_b is approximately equal to zero.
179: The definition of approximately changes as we approach the solution
180: to the problem.
182: No one rule is guaranteed to work in all cases. The following
183: definition is based on the norm of the Jacobian matrix. If the
184: norm is large, the tolerance becomes smaller. */
185: MatNorm(tao->jacobian,NORM_1,&asls->identifier);
186: asls->identifier = PetscMin(asls->merit, 1e-2) / (1 + asls->identifier);
188: VecSet(asls->t1,-asls->identifier);
189: VecSet(asls->t2, asls->identifier);
191: ISDestroy(&asls->fixed);
192: ISDestroy(&asls->free);
193: VecWhichBetweenOrEqual(asls->t1, asls->db, asls->t2, &asls->fixed);
194: ISComplementVec(asls->fixed,asls->t1, &asls->free);
196: ISGetSize(asls->fixed,&nf);
197: PetscInfo1(tao,"Number of fixed variables: %D\n", nf);
199: /* We now have our partition. Now calculate the direction in the
200: fixed variable space. */
201: TaoVecGetSubVec(asls->ff, asls->fixed, tao->subset_type, 0.0, &asls->r1);
202: TaoVecGetSubVec(asls->da, asls->fixed, tao->subset_type, 1.0, &asls->r2);
203: VecPointwiseDivide(asls->r1,asls->r1,asls->r2);
204: VecSet(tao->stepdirection,0.0);
205: VecISAXPY(tao->stepdirection, asls->fixed, 1.0,asls->r1);
207: /* Our direction in the Fixed Variable Set is fixed. Calculate the
208: information needed for the step in the Free Variable Set. To
209: do this, we need to know the diagonal perturbation and the
210: right hand side. */
212: TaoVecGetSubVec(asls->da, asls->free, tao->subset_type, 0.0, &asls->r1);
213: TaoVecGetSubVec(asls->ff, asls->free, tao->subset_type, 0.0, &asls->r2);
214: TaoVecGetSubVec(asls->db, asls->free, tao->subset_type, 1.0, &asls->r3);
215: VecPointwiseDivide(asls->r1,asls->r1, asls->r3);
216: VecPointwiseDivide(asls->r2,asls->r2, asls->r3);
218: /* r1 is the diagonal perturbation
219: r2 is the right hand side
220: r3 is no longer needed
222: Now need to modify r2 for our direction choice in the fixed
223: variable set: calculate t1 = J*d, take the reduced vector
224: of t1 and modify r2. */
226: MatMult(tao->jacobian, tao->stepdirection, asls->t1);
227: TaoVecGetSubVec(asls->t1,asls->free,tao->subset_type,0.0,&asls->r3);
228: VecAXPY(asls->r2, -1.0, asls->r3);
230: /* Calculate the reduced problem matrix and the direction */
231: TaoMatGetSubMat(tao->jacobian, asls->free, asls->w, tao->subset_type,&asls->J_sub);
232: if (tao->jacobian != tao->jacobian_pre) {
233: TaoMatGetSubMat(tao->jacobian_pre, asls->free, asls->w, tao->subset_type, &asls->Jpre_sub);
234: } else {
235: MatDestroy(&asls->Jpre_sub);
236: asls->Jpre_sub = asls->J_sub;
237: PetscObjectReference((PetscObject)(asls->Jpre_sub));
238: }
239: MatDiagonalSet(asls->J_sub, asls->r1,ADD_VALUES);
240: TaoVecGetSubVec(tao->stepdirection, asls->free, tao->subset_type, 0.0, &asls->dxfree);
241: VecSet(asls->dxfree, 0.0);
243: /* Calculate the reduced direction. (Really negative of Newton
244: direction. Therefore, rest of the code uses -d.) */
245: KSPReset(tao->ksp);
246: KSPSetOperators(tao->ksp, asls->J_sub, asls->Jpre_sub);
247: KSPSolve(tao->ksp, asls->r2, asls->dxfree);
248: KSPGetIterationNumber(tao->ksp,&tao->ksp_its);
249: tao->ksp_tot_its+=tao->ksp_its;
251: /* Add the direction in the free variables back into the real direction. */
252: VecISAXPY(tao->stepdirection, asls->free, 1.0,asls->dxfree);
255: /* Check the projected real direction for descent and if not, use the negative
256: gradient direction. */
257: VecCopy(tao->stepdirection, asls->w);
258: VecScale(asls->w, -1.0);
259: VecBoundGradientProjection(asls->w, tao->solution, tao->XL, tao->XU, asls->w);
260: VecNorm(asls->w, NORM_2, &normd);
261: VecDot(asls->w, asls->dpsi, &innerd);
263: if (innerd >= -asls->delta*PetscPowReal(normd, asls->rho)) {
264: PetscInfo1(tao,"Gradient direction: %5.4e.\n", (double)innerd);
265: PetscInfo1(tao, "Iteration %D: newton direction not descent\n", tao->niter);
266: VecCopy(asls->dpsi, tao->stepdirection);
267: VecDot(asls->dpsi, tao->stepdirection, &innerd);
268: }
270: VecScale(tao->stepdirection, -1.0);
271: innerd = -innerd;
273: /* We now have a correct descent direction. Apply a linesearch to
274: find the new iterate. */
275: TaoLineSearchSetInitialStepLength(tao->linesearch, 1.0);
276: TaoLineSearchApply(tao->linesearch, tao->solution, &psi,asls->dpsi, tao->stepdirection, &t, &ls_reason);
277: VecNorm(asls->dpsi, NORM_2, &ndpsi);
278: }
279: return(0);
280: }
282: /* ---------------------------------------------------------- */
283: /*MC
284: TAOASFLS - Active-set feasible linesearch algorithm for solving
285: complementarity constraints
287: Options Database Keys:
288: + -tao_ssls_delta - descent test fraction
289: - -tao_ssls_rho - descent test power
291: Level: beginner
292: M*/
293: PETSC_EXTERN PetscErrorCode TaoCreate_ASFLS(Tao tao)294: {
295: TAO_SSLS *asls;
297: const char *armijo_type = TAOLINESEARCHARMIJO;
300: PetscNewLog(tao,&asls);
301: tao->data = (void*)asls;
302: tao->ops->solve = TaoSolve_ASFLS;
303: tao->ops->setup = TaoSetUp_ASFLS;
304: tao->ops->view = TaoView_SSLS;
305: tao->ops->setfromoptions = TaoSetFromOptions_SSLS;
306: tao->ops->destroy = TaoDestroy_ASFLS;
307: tao->subset_type = TAO_SUBSET_SUBVEC;
308: asls->delta = 1e-10;
309: asls->rho = 2.1;
310: asls->fixed = NULL;
311: asls->free = NULL;
312: asls->J_sub = NULL;
313: asls->Jpre_sub = NULL;
314: asls->w = NULL;
315: asls->r1 = NULL;
316: asls->r2 = NULL;
317: asls->r3 = NULL;
318: asls->t1 = NULL;
319: asls->t2 = NULL;
320: asls->dxfree = NULL;
321: asls->identifier = 1e-5;
323: TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch);
324: PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);
325: TaoLineSearchSetType(tao->linesearch, armijo_type);
326: TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);
327: TaoLineSearchSetFromOptions(tao->linesearch);
329: KSPCreate(((PetscObject)tao)->comm, &tao->ksp);
330: PetscObjectIncrementTabLevel((PetscObject)tao->ksp, (PetscObject)tao, 1);
331: KSPSetOptionsPrefix(tao->ksp,tao->hdr.prefix);
332: KSPSetFromOptions(tao->ksp);
334: /* Override default settings (unless already changed) */
335: if (!tao->max_it_changed) tao->max_it = 2000;
336: if (!tao->max_funcs_changed) tao->max_funcs = 4000;
337: if (!tao->gttol_changed) tao->gttol = 0;
338: if (!tao->grtol_changed) tao->grtol = 0;
339: #if defined(PETSC_USE_REAL_SINGLE)
340: if (!tao->gatol_changed) tao->gatol = 1.0e-6;
341: if (!tao->fmin_changed) tao->fmin = 1.0e-4;
342: #else
343: if (!tao->gatol_changed) tao->gatol = 1.0e-16;
344: if (!tao->fmin_changed) tao->fmin = 1.0e-8;
345: #endif
346: return(0);
347: }