A CONIC INTERIOR POINT DECOMPOSITION APPROACH FOR LARGE SCALE SEMIDEFINITE PROGRAMMING

Size: px

Start display at page:

Download "A CONIC INTERIOR POINT DECOMPOSITION APPROACH FOR LARGE SCALE SEMIDEFINITE PROGRAMMING"

Bennett Grant
6 years ago
Views:

1 A CONIC INTERIOR POINT DECOMPOSITION APPROACH FOR LARGE SCALE SEMIDEFINITE PROGRAMMING Kartik Krishnan Sivaramakrishnan Department of Mathematics NC State University kksivara Operations Research Colloquia University of North Carolina at Chapel Hill April 12, 2006

2 Contents Conic programming Motivation for our decomposition approach A decomposition framework for conic programming Semidefinite programming Conic Dantzig-Wolfe decomposition Algorithm Implementational Issues in Algorithm Computational results Conclusions -1-

3 Conic programming (P) min c T x s.t. Ax = b x K (D) max b T y s.t. A T y + s = c s K where A R m n, b R m, c R n, K = K 1... K r r = 1, K = R n + = {x R n : x 0} LP Very large LPs (m, n 1,000,000) solvable by the simplex method and/or IPMs. K i = Q n i + = {x R n i : x 1 x 2:ni } SOCP Large SOCPs (m, n 100,000) solvable by IPMs. K i = S n i + = {X S n i : X 0} SDP Medium sized SDPs (m, n 1000) solvable by IPMs. (Beyond 10,000 seems impossible today!) -2-

4 Motivation (a) Solve large scale structured semidefinite programs (SDP) arising in science and engineering. Typically, these SDPs need not be solved very accurately. (b) The technique is to iteratively solve an SDP between a mixed conic master problem over linear, second order, and semidefinite cones; and distributed subproblems (smaller SDPs) in a high performance computing environment. (c) Improve the scalability of interior point methods (IPMs) by applying them instead on the smaller master problem, and subproblems (which are solved in parallel!) This is our conic interior point decomposition scheme. -3-

5 MIN sum(i=1)^k f_i(x_i) s.t. sum(i=1)^k A_i(X_i) = b COUPLING CONSTRAINTS X_i in C_i, i=1,...,k SUBPROBLEMS MASTER PROBLEM MAX theta(y) where theta(y) = b^ty + sum(i=1)^k theta_i(y) PROPOSALS X_i (PRIVATE VARIABLES) PRICES y (SHARED VARIABLES) theta_1(y) = MIN f_1(x_1) y^ta_1(x_1) s.t. X_1 in C_1 SUBSYSTEM 1 SUBSYSTEM k theta_k(y) = MIN f_k(x_k) y^ta_k(x_k) s.t. X_k in C_k C_i are convex sets defined by LMIs Figure 1: Decomposition by prices

6 Semidefinite programs with a decomposable structure 1. Preprocessed SDPs after matrix completion: Fukuda-Kojima-Nakata-Murota (2000), Lu-Nemirovskii-Monteiro (2004) 2. Stability analysis of interconnected subsystems: Langbort-D Andrea-Xiao-Boyd (2003) 3. Stochastic semidefinite programs: Mehrotra-Ozëvin (2005) 4. Polynomial optimization problems with structured sparsity: Waki-Kim-Kojima-Muramatsu (2005) 5. Semidefinite programs arising in lift and project schemes: Rendl-Sotirov (2003) and Burer-Vandenbussche (2006) -5-

7 Semidefinite programming (SDP) min C X s.t. A(X) = b X 0 (SDD) max b T y s.t. A T y + S = C S 0 Notation - X, S, C S n, b R m - A B = trace(ab) = n i,j=1 A ijb ij (Frobenius inner product) - The operator A : S n R m and its adjoint A T : R m S n are A(X) = A 1 X m., A y = y i A i A m X i=1 where A i S n, i = 1,..., m -6-

8 Assumptions (1) A 1,..., A m are linearly independent in S n. (2) Slater condition for (SDP) and (SDD): {X S n : A(X) = b, X 0} {(y, S) R m S n : A T y + S = C, S 0} (3) One of the constraints in (SDP) is I X = 1 (trace(x) = 1). (Every SDP with a bounded primal feasible set can be reformulated to satisfy this assumption) -7-

9 One sample SDP 1. Consider the following integer quadratic program min x T Lx s.t. x i { 1, 1} n, i = 1,..., n 2. Setting X = xx T, gives an equivalent formulation min L X s.t. X ii = 1, i = 1,..., n X 0 rank(x) = 1 3. Dropping the rank constraint gives an SDP relaxation min L X s.t. X ii = 1, i = 1,..., n X 0 4. Note trace(x) = n in this example. -8-

10 Conic Dantzig-Wolfe decomposition Our semidefinite programs are: Consider the set: The extreme points of E are: (SDP) min C X s.t. A(X) = b trace(x) = 1 X 0 (SDD) max b T y + z s.t. A T y + zi + S = C S 0 E = {X S n : trace(x) = 1, X 0} {vv T : v R n, v T v = 1} (Infinite number) -9-

11 Conic Dantzig-Wolfe decomposition: Primal motivation Any X E can be written as: (1) X = j λ jd j d T j where j λ j = 1, λ j 0 X 0 is replaced by λ j 0, j (Semi-infinite LP formulation) (2) X = j P jv j P T j where j trace(v j) = 1, V j S 2 + X 0 is replaced by V j 0, j (Semi-infinite SOCP formulation) (A 2 2 SDP constraint can be written as a size 3 SOCP constraint) (3) X = j P jv j P T j where j trace(v j) = 1, V j S r j +, r j 3 X 0 is replaced by V j 0, j (Semi-infinite SDP formulation) -10-

12 A 2 2 SDP constraint is an SOCP constraint of size 3 X = ( ) X11 X 12 X 12 X 22 0 X 11 0, X 22 0, X 11 X 22 X X 11 + X 22 2X 12 X 11 X 22 Q 0-11-

13 Conic Dantzig-Wolfe decomposition: Dual motivation (1) LP formulation: S 0 d T j Sd j 0, j (if we take a finite number of constraints we obtain an LP relaxation) (2) SOCP formulation: S 0 P T j SP j }{{} 2 2 0, j (finite number of constraints gives SOCP relaxation) (3) SDP formulation: S 0 P T j SP j }{{} 0, j (r j 3, j) r j r j (finite number of constraints gives SDP relaxation) -12-

14 Conic Dantzig-Wolfe decomposition: Master problem The decomposed conic problem over LP, SOCP and SDP blocks is: Primal Dual min s.t. n l c li x li + c T qjx qj + C sk X sk i=1 n l n q j=1 n s k=1 A li x li + A qj x qj + A sk (X sk ) = b i=1 n l n q j=1 n s k=1 x li + x qj1 + trace(x sk ) = 1 i=1 n q j=1 n s k=1 x li 0, i = 1,..., n l x qj Q 0, j = 1,..., n q X sk 0, k = 1,..., n s max b T y + z s.t. A T li y + z + s li = 1 c li, i = 1,..., n l A T qj y + z 0 + s qj 0 = c qj, j = 1,..., n q A T sk y + zi r k + S sk = C sk, k = 1,..., n s s li 0, i = 1,..., n l s qj Q 0, j = 1,..., n q S sk 0, k = 1,..., n s -13-

15 Conic Dantzig-Wolfe decomposition: Separation Oracle Input: (y, z ) a feasible point for dual master problem. If λ min (S ) 0, report feasibility STOP. Else, solve the following subproblem for X : min (C A T y z I) X s.t. I X = 1 X 0 Factorize X = DMD T with M 0. Return the cut D T (C A T y zi)d 0 with D R n r. If r = 1 is an LP cut. If r = 2 is an SOCP cut. If r 3 is an SDP cut of small size. -14-

16 Figure 2: Conic Dantzig-Wolfe Algorithm First Prev Next Last Go Back Full Screen Close Quit

17 Choice of the query point: I In the original Dantzig-Wolfe (Kelley) scheme the query point (y, z) is an optimal solution to the dual master problem. This scheme has a poor rate of convergence. Better to solve the master problem approximately initially, and gradually tighten this tolerance as the algorithm proceeds. Initially, the master problem is a poor approximation to the SDP problem. Weak tolerance Central dual iterates (y, z) Oracle returns better cuts As the algorithm proceeds, the master problem approximations get tighter. Tight tolerance More emphasis on the objective function Convergence to an optimal solution -16-

18 Our adaptive strategy: Choice of the query point : 2 Solve the master problem to a tolerance TOL for (x, y, z, s ). Compute the following parameters: GAPTOL(x, s ) = x T s max{1, 1 2 (ct x +b T y )} λ min (C A T y z I) max{1, 1 2 (ct x +b T y )} INF(y, z ) = OPT(x, y, z, s ) = max{gaptol(x, s ), INF(y, z )} If OPT(x, y, z, s ) < TOL, we lower TOL by a constant factor. More precisely, TOL = µ OPT(x, y, z, s ) with 0 < µ < 1. Else, TOL remains unchanged. -17-

19 Figure 3: The complete algorithm First Prev Next Last Go Back Full Screen Close Quit

20 (y,z) space y New feasible region (SDD) feasible region A D B z New (y,z) Cutting plane C A Central path Old (y,z) Feasible region of dual master problem A : Solve master problem to tolerance TOL B : Oracle returns a cutting plane C : Restoring dual feasibility D : Warm start Figure 4: One iteration of algorithm First Prev Next Last Go Back Full Screen Close Quit

21 Upper and lower bounds The objective value of the primal master problem in every iteration gives an upper bound on the SDP objective value. This is the objective value of the dual master problem plus the duality gap. Given a solution (y, z ) to the dual master problem in every iteration, a lower bound is computed as follows: Compute λ = λ min (S ), where S = (C A T y z I). Set y lb = y and z lb = z + λ. A lower bound on the SDP objective value is then b t y lb + z lb Note: (y lb, z lb ) is a feasible point in (SDD). We could also terminate if the difference between these bounds is small. -20-

22 Warm start after adding a column: 1 The solution to the old master problem is (x l, x s, y, s l, s s). Consider one linear cut a T l y d in the dual master problem. The new master problem is : min c T l x l + c T s x s + d T β s.t. A l x l + A s x s + a l β = b x l 0, β 0 x s 0. max b T y s.t. s l = c l A T l y 0 s s = c s A T s y Q 0 γ = d a T l y 0 We have γ < 0. We can perturb y (lower bounding) to generate y lb so that γ lb = 0. The perturbed point (x l, x s, β, y lb, s lb l, s lb s, γ lb ) is feasible in the new master problem with β = γ lb = 0, but NOT STRICTLY!. We want to increase β and γ lb from their current zero values, while limiting the variation in the other variables. -21-

23 Warm start after adding a column: 2 We solve the following problems (WSP) max log β s.t. A l x l + A s x s + a l β = 0 D 1 l x l 2 + Ds 1 x s 2 1 β 0. (WSD) max log γ s.t. a T l y + γ = 0 Dl A T l y 2 + D s A T s y 2 1 γ 0. for ( x l, x s, β) and ( y, γ) respectively. Here D l and D s are appropriate primal-dual scaling matrices for linear and semidefinite blocks at (x l, x s, s lb l, s lb s ). Compute ( s l, s s ) = ( A T l y, A T s y). -22-

24 Warm start after adding a column: 3 The solutions to (WSP) and (WSD) are given by x l = D 2 l A T l (A l D 2 l A T l + A s D 2 sa T s ) 1 a l β x s = D 2 qa T q (A l D 2 l A T l + A s D 2 sa T s ) 1 a l β y = (A l D 2 l A T l + A s D 2 sa T s ) 1 a l β γ = V β where β R is the solution to min{ 1 2 βt V β log β : β 0} where V = a T l (A l D 2 l A T l + A s D 2 sa T s ) 1 a l. Finally set (x st l, x st s, β st ) = (x l + καmax x p l, x s + καmax x p s, καmaxβ) p y st = yl lb + καmax y d + καmax s d l, s lb s + καmax s d s, καmaxγ) d (s st l, s st s, γ st ) = (s lb l where κ (0, 1) and α p max, α d max are the maximal primal and dual step lengths respectively. -23-

25 Warm start after adding a matrix: 1 The solution to the old master problem is (x l, x s, y, s l, s s). Consider adding a semidefinite cut a T s y d in the dual master problem. The new master problem is : min c T l x l + c T s x s + d T β s.t. A l x l + A s x s + a s β = b x l 0, β 0 x s 0. max b T y s.t. s l = c l A T l y 0 s s = c s A T s y 0 γ = d a T s y 0 We have γ 0. We can perturb y (lower bounding) to generate y lb so that γ lb 0. The perturbed point (x l, x s, β, y lb, s lb l, s lb s, γ lb ) is feasible in the new master problem with β 0 and γ lb 0, but NOT STRICTLY!. We want to increase β and γ lb, making them strictly feasible, while limiting the variation in the other variables. -24-

26 Warm start after adding a matrix: 2 We solve the following problems (WSP) max log det β s.t. A l x l + A s x s + a s β = 0 D 1 l x l 2 + Ds 1 x s 2 1 β 0. (WSD) max log det γ s.t. a T s y + γ = 0 Dl A T l y 2 + D s A T s y 2 1 γ 0. for ( x l, x s, β) and ( y, γ) respectively. Here D l and D s are appropriate primal-dual scaling matrices for LP and SDP blocks at (x l, x s, s lb l, s lb s ). Compute ( s l, s s ) = ( A T l y, A T s y). -25-

27 Warm start after adding a matrix: 3 Compute x l = D 2 l A T l (A l D 2 l A T l + A s D 2 sa T s ) 1 a s β x s = D 2 sa T q (A l D 2 l A T l + A s D 2 sa T s ) 1 a s β y = (A l D 2 l A T l + A s D 2 sa T s ) 1 a s β γ = V β where β S p is the solution to min{ p 2 βt V β log det β : β 0} where V = a T s (A l D 2 l A T l + A s D 2 sa T s ) 1 a s. Finally set (x st l, x st s, β st ) = (x l + καmax x p l, x s + καmax x p s, καmaxβ) p y st = yl lb + καmax y d + καmax s d l, s lb s + καmax s d s, καmaxγ) d (s st l, s st s, γ st ) = (s lb l where κ (0, 1) and α p max, α d max are the maximal primal and dual step lengths respectively. -26-

28 Aggregating unimportant blocks Use a primal-dual scaling measure to evaluate blocks: For the ith linear block the measure is x i s i. For the jth semidefinite block the measure is trace(x j) trace(s j ). Blocks with small measure are aggregated in an LP block. The resulting master problem is min c agg x agg + c T l x l + c T s x s s.t. A agg x agg + A l x l + A s x s = b x l 0, x agg 0 x s 0. max b T y s.t. s l = c l A T l y 0 s s = c s A T s y 0 s agg = c agg A T aggy 0-27-

29 Computational Results Matlab code with the following features: Master problem contains linear and semidefinite blocks only. SDPT3 is used as conic solver for master problem with H.K.M. scaling Oracle is implemented using MATLAB s Lanczos solver (eigs), which computes the r = m most negative eigenvalues and eigenvectors 2 of S in each iteration. Solve the problem for m iterations, or until TOL = 1e 3, whichever comes earlier. All computational results on a 2.4 GHz processor with 1 GB of memory. -28-

30 Computational results for Max-Cut Problem n Opt LP SDP Lower ɛ Time value cones cones bound (h:m:s) mcp (31) e-3 16 mcp (48) e-3 34 mcp (36) e-3 25 mcp (35) e-3 22 mcp (41) e-3 21 mcp (84) e-3 2:03 mcp (51) e-3 1:05 mcp (44) e-3 49 mcp (43) e-3 47 mcp (115) e-3 12:29 mcp (70) e-3 6:18 mcp (57) e-3 4:33 mcp (49) e-3 2:51 toruspm (84) e-3 9:28 torusg (89) e maxg (84) e-3 1:49:55 maxg (54) e-3 32:24 First Prev Next Last Go Back Full Screen Close Quit

31 0 100 Lower Bound Upper Bound Optimal Value Objective Value for maxg Iteration Number Figure 5: Variation of bounds for maxg11 problem

32 1 0 1 log(epsilon) Iteration Number Figure 6: Variation of error for maxg11 problem

33 Rest of code 3% 13% Subproblem 20% Warm Start 64% Master Problem Figure 7: Distribution of times for maxg11 problem

34 Computational results for graph partitioning Problem n Opt LP SDP Lower ɛ Time value cones cones bound (h:m:s) gpp (23) e-3 50 gpp (52) e-2 1:35 gpp (39) e-3 1:24 gpp (42) e-3 50 gpp (39) e-3 43 gpp (66) e-2 10:24 gpp (49) e-3 10:11 gpp (50) e-3 3:55 gpp (53) e-3 3:49 gpp (92) e-2 1:48:23 gpp (65) e-3 2:11:08 gpp (63) e-3 40:21 gpp (56) e-3 26:22 bm (339) e-2 26:31:13-33-

35 Computational results for Lovasz theta Problem n m Opt LP SDP Lower ɛ Time value blocks blocks bound (h:m:s) theta (35) e-3 1:09 theta (37) e-3 38:20 theta (40) e-3 5:15:52-34-

36 Computational results on random semidefinite programs max b T y + z s.t. S = (C A T y zi) 0 l y u. Prob n m Den LP SDP Our Our SDPT3 SDPT3 cones cones LB (h:m:s) LB (h:m:s) test (9) : test (28) : :32 test (25) : :31 test (29) : :35 test (12) : :03 test (19) : test (21) : :13 test (18) : :11-35-

37 Ongoing and Future work Applying the conic interior point decomposition approach for structured block angular SDPs in a parallel computing environment. (an ongoing project with Brian Borchers at New Mexico Tech) Incorporating the decomposition scheme in the pricing phase of an SDP based conic branch-cut-price scheme for mixed integer and non-convex problems. -36-

38 Consider the SDP Preprocessing sparse SDPs into block-angular SDPs: 1 min L X s.t. X ii = 1, i = 1,..., n, X 0, where L is the adjacency matrix of the graph C2 = (2,3,4,6) C2 C4 = (5,6,7) C4 C3 C1 = (1,5,7) C1 C3 = (4,6,7) GRAPH CHORDAL EXTENSION OF GRAPH -37-

39 Preprocessing sparse SDPs into block-angular SDPs: 2 Using matrix completion, one can reformulate the earlier SDP as min 4 (L k X k ) k=1 s.t. X23 1 X13 4 = 0, X34 2 X12 3 = 0, X23 3 X23 4 = 0, Xii k = 1, i = 1,..., C k, k = 1,..., 4, X k 0, k = 1,..., 4, which is in block-angular form. -38-

40 Thank you for your attention!. Questions, Comments, Suggestions? The slides from this talk are available online at kksivara/ publications/unc-or.pdf A technical report appears at kksivara/ publications/conic-ipm-decomposition.pdf -39-

41 Bibliography: 1 1. K. Sivaramakrishnan, A conic interior point decomposition for large scale structured block angular semidefinite programs, forthcoming. 2. K. Sivaramakrishnan, G. Plaza, and T. Terlaky, A conic interior point decomposition approach for semidefinite programming, submitted. kksivara/publications/conic-ipm-decomposition.pdf 3. L. Lasdon, Optimization Theory for Large Systems, Dover Publications Inc., A. Ruszczyński, Nonlinear Optimization, Princeton University Press, J.B. Hiriart-Urruty and C. Lemaréchal, Convex Analysis and Minimization Algorithms, Springer Verlag, New York, R.H. Tütüncü, K.C. Toh, and M.J. Todd, Solving semidefinite-quadratic-linear programs using SDPT3, Mathematical Programming, 95(2003), pp mattohkc/sdpt3.html 7. B. Borchers, SDPLIB 1.2, A library of semidefinite programming test problems, Optimization Methods and Software, 11, 1999, pp sdplib/ 8. DIMACS 7 th challenge M. Fukuda, M. Kojima, K. Murota, and K. Nakata, Exploiting sparsity in semidefinite programming via matrix completion I: General framework, SIAM Journal on Optimization, 11(2000), pp

42 Bibliography: Z. Lu, A. Nemirovskii, and R.D.C. Monteiro, Large scale semidefinite programming via saddle point mirror-prox algorithm. HTML/2004/11/998.html 11. C. Langbort, R. D Andrea, L. Xiao, and S. Boyd, A decomposition approach to distributed analysis of networked systems, Proceedings of the 43rd IEEE Conference on Decision and Control, pp boyd/distr contr cdc04.html 12. S. Mehrotra and M.G. Özevin, On the implementation of interior point decomposition algorithms for two-stage stochastic conic programs. HTML/2005/10/1237.html 13. H. Waki, S. Kim, M. Kojima, and M. Muramatsu, Sums of squares and semidefinite programming relaxations for polynomial problems with structured sparsity. HTML/2004/10/988.html 14. F. Rendl and R. Sotirov, Bounds for the quadratic assignment problem using the bundle method. HTML/2003/08/712.html 15. S. Burer and D. Vandenbussche, Solving lift-and-project relaxations of binary integer programs, SIAM Journal on Optimization, 16(2006), pp

43 Bibliography: M.R. Oskoorouchi and J.L.Goffin, A matrix generation approach for eigenvalue optimization C. Helmberg and F. Rendl, A spectral bundle method for semidefinite programming, SIAM Journal on Optimization, 10(2000), pp K. Sivaramakrishnan and J.E. Mitchell, A semidefinite programming based polyhedral cut and price algorithm for the maxcut problem, Computational Optimization and Applications, 33(2006), pp

A CONIC DANTZIG-WOLFE DECOMPOSITION APPROACH FOR LARGE SCALE SEMIDEFINITE PROGRAMMING

A CONIC DANTZIG-WOLFE DECOMPOSITION APPROACH FOR LARGE SCALE SEMIDEFINITE PROGRAMMING Kartik Krishnan Advanced Optimization Laboratory McMaster University Joint work with Gema Plaza Martinez and Tamás