Towards a practical simplex method for second order cone programming

Transcription

1 Towards a practical simplex method for second order cone programming Kartik Krishnan Department of Computing and Software McMaster University Joint work with Gábor Pataki (UNC), Neha Gupta (IIT Delhi), and Tamás Terlaky (Mac) INFORMS Annual Meeting Denver Oct 26, 2004

2 Overview Second Order Cone Programming (SOCP) Contrast simplex-like approaches and IPMs for conic optimization The geometry of SOCP Simplex algorithm for SOCP Special case: Simplex method for LP Properties of the algorithm Preliminary computational results Conclusions and future work 1

3 Second order cone optimization Primal min c T 1 x 1 + c T 2 x c T r x r A 1 x 1 + A 2 x A r x r = b, (SOCP) x i K i. Dual max b T y A T i y + s i = c i, i = 1,..., r, s i K i. (SOCD) Notation A = (A 1, A 2,..., A r ) R m n with full row rank. K = K 1... K r. Each K i = { x R n i : x 1 of size n i, i = 1,..., r. x 2:ni } is a second order cone 2

4 Optimality conditions (x, y, s ) are optimal iff Ax = b, x K, A T y + s = c, s K, x i s i = 0, i = 1,..., r. (CS) (PF) (DF) For an SOCP cone in R n x s = ( x T s x(1)s(2 : n) + s(1)x(2 : n) ). For any cone K x K, s K, x T s = 0 x s = 0. 3

5 Eigenvalues and Eigenvectors Given x R n, we have ( x = 1 2 ( x 1 x 2:n ) 1; x ) 2:n + 1 x 2:n 2 ( x 1 + x 2:n ) = λ min ( x)v min ( x) + λ max ( x)v max ( x). Index classification: Given x i K i we have i O (zero blocks) if λ max ( x i ) = 0. i R (boundary blocks) if λ min ( x i ) = 0. i I (interior blocks) if λ min ( x i ) > 0. ( 1; ) x 2:n x 2:n 4

6 Contrast simplex and IPMs for CP I. Start Extreme Points Start Central Path Our Algorithm Simplex Method Finish Interior Point Method Finish Feasible Direction Method Interior Point Methods deal with matrices of full rank. In the simplex method, the rank of the extreme points satisfy (1) r <= m (Linear Programming) (2) r(r+1)/2 <= m (Semidefinite Programming) 5

7 Contrast simplex and IPMs for CP II. Why a simplex method for conic programming?. 1. Listed as an important open problem in conic programming. 2. Warm start after branching or the addition of cutting planes using the dual simplex method. 3. It is possible to do every simplex iteration more quickly than an IPM iteration using fast basis LU updates to factorize the basis matrix. 6

8 Terminology I. Given a closed convex cone K R n. Consider x K Lineality space: B x = {d R n : x ± ǫd K, ǫ > 0}. Tangent space: T x = { d R n : dist( x ± ǫd, K) = O(ǫ 2 ), ǫ > 0 }. Residual space: R x = T x \ B x. 7

9 Terminology II. Cone of feasible directions: Tangent cone: dir( x, K) = (K + B x ) = {d R n : x + ǫd K, ǫ > 0}. TC( x, K) = cl(dir( x, K)) = { d R n : dist( x + ǫd, K) = O(ǫ 2 ), ǫ > 0 }. Null space: N = {x R n : Ax = 0}. 8

10 Demo Given x i R n i on the boundary of the SOCP cone K i. B xi = α x i for α R. T xi = lin ( B xi ( 0 {( 0 w ) : w T x 2:ni = 0 Note x i + ǫ / K w i. ( ) ( 0 However x i + ǫ + ǫ w ) ) }). int(k i ). 9

11 Notions of nondegeneracy A feasible x in (SOCP) is c-nondegenerate This is a generic property. T x + N = R n. f-nondegenerate B x + N = R n. Extreme point B x N = {0}. 10

12 Simplex algorithm for SOCP I. Given a feasible c-nondegenerate x = ( x I ; x R ; x O ) in (SOCP). Indices i I and j R are ranked on λ min ( x i ) and x j1 respectively. 1. Select basis: Construct the following basis elements: Decompose x = M B x B + M N x N, where M B R n m and M N R n (n m). M B is chosen with Range(M B ) T x such that A B = AM B is a nonsingular basis matrix of size m. Also, c B = M T B c. Let B,N be the index set of cones K i,i = 1,..., r with columns in M B and M N respectively. (Note: B N φ). 2. Construct dual solution: Solve A T B y = c B for ȳ. Compute s i = c i A T i ȳi, i N. (Note: s i = 0, i / N). 11

13 Simplex algorithm for SOCP II. 3. Pricing: Compute α = max{ λ min ( s i ) : s i / K i, i N}. β = max{ x T i s i : s i K i, i N}. If α = β = 0 STOP; ( x, ȳ, s) is an optimal solution. Else k = { an index s.t. α = λmin ( s k ) if α > β. an index s.t. β = x T k s k if α < β. 4. Find improving direction: Improving component: Compute d N R n where d Ni = min d =1 d TC( x k,k ) st k d k (IC) if i = k 0 for i = 1,..., r and i k. (Note: dir( x k, K k ) is not CLOSED!) 12

14 Simplex algorithm for SOCP III. 4. Find improving direction (continued): Centering component: Construct d C R n satisfying d Ci = (α;0) i R B, a nonzero subset of R xi Range(A B ) (α;0) i = k if k R 0 otherwise with α > 0 chosen so that s T k ( d N + d C ) < 0. Basic component: Compute d B R m by solving A B d B = (Ad N + Ad C ). The improving direction d = (M B d B + d C + d N ). 13

15 Simplex algorithm for SOCP IV. 5. Line search: Compute ᾱ where ᾱ = max{α i : x i + α i d i K i, i = 1,..., r}. If ᾱ = the primal is unbounded. STOP. Else set x = x + ᾱ d and return to step 1. (Note: The new x is assumed to be c-nondegenerate). 14

16 Computing the improving component I. The solution d Nk to (IC) is If k O: d Nk = s k s k if λ max ( s k ) < 0 v min ( s k ) if λ min ( s k ) < 0 and λ max ( s k ) 0 0 if s k K k If k I: d Nk = s k s k if s k 0 0 if s k = 0 15

17 Computing the improving component II. If k R: We have TC( x k, K k ) = {d R n k : x 1 d 1 n k j=2 x jd j 0} In this case d Nk is d Nk = 0 if s k = ( x 1 ; x 2:nk ) min d =1, d TC( x k,k ) st k d otherwise k 16

18 Special case: Simplex method for LP 1. The initial iterate is an nondegenerate extreme point solution. If this iterate is not an extreme point, the method is instead a feasible direction method. 2. Given a non-degenerate extreme point, the simplex method chooses the basis matrix as A B = A(:,support( x)). This is nonsingular with Range(A B ) = B x. 3. For the improving direction d N = e k (k is the index for which s i is the most negative) and d B the solution to A B d B = A k. No centering term d C is needed. The resulting direction d is along an edge of the feasible set. 17

19 Properties of the algorithm Theorem. Let {(x k, y k, s k )} be a sequence generated by the algorithm. Then for all k, x k is primal feasible. At the k-th iteration, one of the following alternative cases arises: 1. If it stops in Step 2 then x k, (y k, s k ) are primal and dual optimal solutions to (SOCP) and (SOCD) respectively. 2. If it stops in Step 5, then (SOCP) is unbounded and (SOCD) is infeasible. 3. Otherwise, if x k is also c-nondegenerate we have c T x k+1 < c T x k. 18

20 Convergence of the algorithm Some practical issues affecting convergence include: One also adds a centering term d Ci to the cones i I which are very nearly in R. This is to prevent the algorithm from getting jammed at a suboptimal point. There is also the issue of zigzagging with the algorithm. Conjecture. If {x k } contains a nondegenerate subsequence, then the algorithm with the anti-jamming safeguard either terminates in a finite number of steps, or all accumulation points of {x k } are optimal solutions to (SOCP). 19

21 Comments on the simplex algorithm The simplex algorithm for SOCP is a feasible direction method that generates search directions in the tangent cone. This direction is suitably centered so as to generate a feasible direction. The simplex iterates are not always extreme points and the search directions may traverse the interior of the feasible region. The method resembles the convex simplex method of Zangwill for minimizing a convex function over a polyhedron. The algorithm maintains primal feasibility in every iteration while dual feasibility and complementary slackness are attained at optimality. The step length calculation in Step 5 has a closed analytic expression. 20

22 Preliminary computational results Prob m n lp r Opt Obj(0) Obj(iter) Iter hs [12] (*) 10 slp (3) slp (3) slp [50 10] (*) 183 slp (3) (*) 896 slp (3) (*) 153 slp (5) slp (5) slp [10] (*) 157 slp [10] (*) 128 slp [ ] slp [ ]

23 Conclusions and future work 1. A primal simplex approach for conic optimization of which the primal simplex method for LP is a special case. 2. The simplex approach exploits the well known facial structure of SOCP problems developed in Pataki (2000). 3. We have the framework for solving conic optimization problems over LP, SOCP and SDP cones. 4. We also have a dual simplex variant which mimics the dual simplex method for LP. 5. Currently investigating fast basis inverse (LU) updates to speed up the algorithm. 6. Future use in warm start after branching or the addition of cutting planes. 22

24 References F. Alizadeh and D. Goldfarb. Second-order cone programming, Mathematical Programming, 95(2003), pp F. Alizadeh, J.P.A. Haeberly, and M.L. Overton. Complementarity and Nondegeneracy in SDP, Mathematical Programming, 77(1997), pp M.S. Bazaraa, H.D. Sherali and C.M. Shetty. Nonlinear Programming: Theory and Algorithms, 2nd edition, John Wiley K. Krishnan, G. Pataki, N. Gupta and T. Terlaky. A feasible direction method for SOCP, forthcoming. 23

25 I. Maros. Computational techniques of the simplex method. Kluwer International Series, G. Pataki. Cone-LP s and Semidefinite Programs: Geometry and a Simplex-type Method, Proceedings of the 5th IPCO conference, Springer Verlag G. Pataki. The Geometry of Semidefinite Programming, Handbook of SDP edited by H. Wolkowicz et al., Kluwer Academic Publishers, 2000, pp