S-free Sets for Polynomial Optimization

Transcription

1 S-free Sets for and Daniel Bienstock, Chen Chen, Gonzalo Muñoz, IEOR, Columbia University May, 2017 SIAM OPT 2017 S-free sets for PolyOpt 1

2 The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions SIAM OPT 2017 S-free sets for PolyOpt 2

3 The Polyhedral Approach The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions Consider a mathematical program of the following form minc T x subject tox S P. P := {x R n Ax b} is a polyhedral set, ands R n is a closed set. Can we strengthenp with cuts? SIAM OPT 2017 S-free sets for PolyOpt 3

4 The Polyhedral Approach The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions Consider a mathematical program of the following form minc T x subject tox S P. P := {x R n Ax b} is a polyhedral set, ands R n is a closed set. Can we strengthenp with cuts? We shall focus on the geometric approach: cuts vias-free sets. (Many other ways to generate cuts, e.g. disjunctions, algebraic arguments, combinatorics, convex cuts, etc.) SIAM OPT 2017 S-free sets for PolyOpt 3

5 TighteningP with ans-free setc The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions P S LetS be some closed set. Example: a rectanglep. Want to separate the extreme point marked with a red circle froms. SIAM OPT 2017 S-free sets for PolyOpt 4

6 TighteningP with ans-free setc The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions C P S C is ans-free set [Dey and Wolsey 2010]: a closed convex set with an interior that does not intersect withs. SIAM OPT 2017 S-free sets for PolyOpt 4

7 TighteningP with ans-free setc The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions C P S We obtain a cut by subtractingc fromp. SIAM OPT 2017 S-free sets for PolyOpt 4

8 TighteningP with ans-free setc The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions conv(pnint(c)) S Applying the single cut gives us conv(p \C). SIAM OPT 2017 S-free sets for PolyOpt 4

9 conv(p \ int(c)) is tricky The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions More than one cut, possibly an infinite number are needed. Separation is generally NP-Hard, e.g. P a polytope,c a ball models an NP-complete set containment problem. SIAM OPT 2017 S-free sets for PolyOpt 5

10 conv(p \ int(c)) is tricky The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions Balas, 1971 (see also Tuy, 1964): IfP is a simplicial cone then the intersection cut guarantees separation over conv(p \ int(c)) Simplicial cone: n linearly independent linear inequalities Simplicial conic relaxationp P is easily obtained from a basic solution ofp With less ambition we go for conv(p \ int(c)) Intersection cut is described in closed form fast separation of extreme points ofp usingp SIAM OPT 2017 S-free sets for PolyOpt 5

11 Intersection cuts and maximals-free sets The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions BiggerC, deeper cuts. C P C 0 S AnS-free setc is maximal S-free if it is not contained in another S-free set. SIAM OPT 2017 S-free sets for PolyOpt 6

12 (Non-Exhaustive) Additional Literature The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature Our Contributions MaximalS-free sets and minimal valid inequalities: [Basu et al. 2010], [Conforti et al. 2014], [Cornuejols, Wolsey, Yildiz, 2015], [Kilinc-Karzan 2015], etc. Intersection cuts and for mixed-integer conic programs programming: [Atamturk and Narayanan 2010], [Belotti et al., 2013], [Andersen and Jensen, 2013], [Dadush, Dey, Vielma 2011], [Modaresi, Kilinc, Vielma 2015/2016], etc. Intersection cuts for bilevel optimization: [Fischetti, Monaci, Sinni, 2016]. Generalized intersection cut procedures: [Balas and Margot, 2013], [Balas, Kazachkov, Margot 2016] SIAM OPT 2017 S-free sets for PolyOpt 7

13 Our Contributions The Polyhedral Approach TighteningP with an S-free setc conv(p \ int(c)) is tricky Intersection cuts and maximals-free sets (Non-Exhaustive) Additional Literature 1) A simple, generic way to generates-free sets that ensures separation. Also, a corresponding cutting plane method for arbitrary closed sets, guaranteed to converge on bounded problems. 2) A study of maximals-free sets for polynomial optimization Our Contributions SIAM OPT 2017 S-free sets for PolyOpt 8

14 Distance Oracle Cut Closure Convergence SIAM OPT 2017 S-free sets for PolyOpt 9

15 Distance Oracle Distance Oracle Cut Closure Convergence Suppose we have an oracle for a closed sets that gives us the distanced(x,s) from any pointx R n to the nearest point ins. SIAM OPT 2017 S-free sets for PolyOpt 10

16 Distance Oracle Distance Oracle Cut Closure Convergence Suppose we have an oracle for a closed sets that gives us the distanced(x,s) from any pointx R n to the nearest point ins. Examples: Integer programming: ifs is the lattice, then one can round. optimization: distance can be calculated to arbitrary accuracy in polynomial time Cardinality constraints: nearest vector of card(k) can be obtained by rounding. SIAM OPT 2017 S-free sets for PolyOpt 10

17 Distance Oracle Distance Oracle Cut Closure Convergence Suppose we have an oracle for a closed sets that gives us the distanced(x,s) from any pointx R n to the nearest point ins. Examples: Integer programming: ifs is the lattice, then one can round. optimization: distance can be calculated to arbitrary accuracy in polynomial time Cardinality constraints: nearest vector of card(k) can be obtained by rounding. Observation. The ball centered aroundxwith radiusd(x,s) is S-free. Call it B(x, d(x, S)). We shall call the corresponding intersection cut an oracle ball cut. SIAM OPT 2017 S-free sets for PolyOpt 10

18 Cut Closure Convergence Distance Oracle Cut Closure Convergence Start with a polytopep 0. DefineP k+1 asp k intersected with conv(p k \ int(b(x,d(x,s)))) for every extreme pointx. This is the rankk oracle cut closure. SIAM OPT 2017 S-free sets for PolyOpt 11

19 Cut Closure Convergence Distance Oracle Cut Closure Convergence Start with a polytopep 0. DefineP k+1 asp k intersected with conv(p k \ int(b(x,d(x,s)))) for every extreme pointx. This is the rankk oracle cut closure. Theorem: lim k P k = conv(s P). SIAM OPT 2017 S-free sets for PolyOpt 11

20 Cut Closure Convergence Distance Oracle Cut Closure Convergence Start with a polytopep 0. DefineP k+1 asp k intersected with conv(p k \ int(b(x,d(x,s)))) for every extreme pointx. This is the rankk oracle cut closure. Theorem: lim k P k = conv(s P). Corollary: given an inexact but arbitrarily accurate distance oracle, we can obtain arbitrarily close (in terms of Hausdorff distance) polyhedral approximation to conv(s P) in finite time. Borrows from proof technique used in [Averkov 2011]. SIAM OPT 2017 S-free sets for PolyOpt 11

21 Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 12

22 Cuts for Lifted Representation S-free sets for z := inf x S p 0(x) S := {x R n p 1 (x) 0,...,p m (x) 0} Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 13

23 Cuts for Cuts for Lifted Representation S-free sets for [Saxena, Bonami, Lee 2010/2011] Disjunctive cuts from MILP inner-approximation + convex cuts Applies to bounded polynomial optimization [Ghaddar, Vera, Anjos 2011] Projections of moment relaxations. Generalizes Balas, Ceria, Cornuejols lifting. Separation not guaranteed in general. Older literature on convex envelopes of functions, e.g. multilinear. Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 14

24 Cuts for Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i [Saxena, Bonami, Lee 2010/2011] Disjunctive cuts from MILP inner-approximation + convex cuts Applies to bounded polynomial optimization [Ghaddar, Vera, Anjos 2011] Projections of moment relaxations. Generalizes Balas, Ceria, Cornuejols lifting. Separation not guaranteed in general. Older literature on convex envelopes of functions, e.g. multilinear. Our intersection cuts (using e.g. the ball) guarantee polynomial-time separation without boundedness assumptions. The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 14

25 Lifted Representation Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 [Shor 1987], [Lovasz and Schrijver 1991] Define a vector of monomials,m = [1,x 1,...,x n,x 1 x 2,x 1 x 3,...,x k n]. Let M = mm T. optimization can be formulated as min A 0,M s.t. A i,m b i,i = 1,...,m. This is a linear programming relaxation with respect tom. A i,m := a ij m ij is the inner product. Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 15

26 Lifted Representation Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 15 [Shor 1987], [Lovasz and Schrijver 1991] Define a vector of monomials,m = [1,x 1,...,x n,x 1 x 2,x 1 x 3,...,x k n]. Let M = mm T. optimization can be formulated as min A 0,M s.t. A i,m b i,i = 1,...,m. This is a linear programming relaxation with respect tom. A i,m := a ij m ij is the inner product. Equivalency when M 0, rank(m) = 1 and consistency constraints. Dropping the rank constraint gives the moment relaxation [Lasserre, SIAMOPT 2001].

27 S-free sets for Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i Geometric notions are with respect to a vectorized space, e.g. M S 2 2 {M 11,M 12,M 22 } R 3 A convex set (in the appropriate vectorized space) is outer-product-free (OPF) if no point in the interior corresponds to a matrix that can be represented as an outer-product. A set is maximal OPF if no OPF set strictly contains it. The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 16

28 S-free sets for Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i Geometric notions are with respect to a vectorized space, e.g. M S 2 2 {M 11,M 12,M 22 } R 3 A convex set (in the appropriate vectorized space) is outer-product-free (OPF) if no point in the interior corresponds to a matrix that can be represented as an outer-product. A set is maximal OPF if no OPF set strictly contains it. It turns out OPF sets can have nice structure and lead to easily generated intersection cuts. The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 16

29 Violation Ball Cuts for Lifted Representation S-free sets for Using a modification by Dax (2016) of the Eckart-Young-Mirksy theorem, we can find the nearest (by Frobenius norm) outer-product to a given matrix. This uses eigenvalues, and hence can be generated (to specified precision) in polynomial time. Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 17

30 Max Full-Dim OPF Sets are Cones Cuts for Lifted Representation S-free sets for Theorem: LetC S n n be an outer-product-free set with full dimension. Then clcone(c) is outer-product-free. Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 18

31 Negative SemidefiniteP i Cuts for Lifted Representation S-free sets for P i,m 0 Theorem: Every such halfspace withp i NSD is maximal outer-product-free. These halfspaces yield the standard outer approximation cut for SDP:M 0 c T Mc 0 c R n. Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 19

32 The2 2cone Cuts for Lifted Representation S-free sets for DefineC i,j := {M S n n M [i,j] 0} Theorem: C i,j is max OPF for1 i j n. Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 20

33 Characterization for2 2 Cuts for Lifted Representation S-free sets for The 3-dimensional case. Theorem: Forn = 2, the maximal OPF set arec 1,2 and halfspaces of the form P,M 0, wherep is NSD. i.e. the PSD cone and halfspaces with boundaries that support the cone Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 21

34 Cuts Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i Recall: we want every extreme point of our relaxation to be an outer-product, i.e. PSD and rank one. Any non-psd extreme point can be separated by the outer-approximation cutsc T Mc 0. Any PSD extreme point with rank greater than 1 can be separated by the intersection cut given byc i,j for somei,j [Chen, Oren, Atamturk 2016]. The oracle cut can be strengthened by recentering the ball and taking the conic hull (complicated expression but computationally fast). The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 22

35 Experiments: Setup Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i Pure cutting plane algorithm implemented in Python using combinations of the following cuts: OB (Oracle Ball Cuts), SO (Strengthened Oracle Cuts) OA (Outer Approximation Cuts), 2x2 (2x2 Principal Minor Cuts) LP solver: Gurobi Hardware: 20-core server, Intel Xeon 3.10GHz CPU, 264 GB RAM 26 QCQP problems from GLOBALLib (6-63 variables) 99 BoxQP instances ( variables) The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 23

36 Experiments: Results Cuts for Lifted Representation S-free sets for Cut Family Initial Gap End Gap Closed Gap # Cuts Iters Time (s) LPTime (%) OB % % 1.00% % SO % 8.77% % OA % 8.61% % 2x2 + OA % 32.61% % SO+2x2+OA % 31.91% % Table 1: Averages for GLOBALLib instances Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 24 Cut Family Initial Gap End Gap Closed Gap # Cuts Iters Time (s) LPTime (%) OB % % 0.04% % SO % 0.34% % OA 30.88% 75.55% % 2x2 + OA 32.84% 74.52% % SO+2x2+OA 33.43% 74.03% % Table 2: Averages for BoxQP instances

37 BoxQP Cuts for Lifted Representation S-free sets for V2: second-order conic outer-approximation of PSD constraint MIP to derive disjunctive cuts Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 25

38 GLOBALLib Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 26 Instance V2 Gap V2 Time Gap Closed Time Ex % % 0.41 Ex % % 0.13 Ex % % 0.95 Ex % % Ex % % 36.9 Ex % % 0.55 Ex % % 0.04 Ex % % 0.26 Ex5 2 2 case1 0.00% % 0.47 Ex5 2 2 case2 0.00% % 0.26 Ex5 2 2 case3 0.36% % 0.16 Ex % % 5.69 Table 3: Comparison with V2 on GLOBALLib instances

39 GLOBALLib Cuts for Lifted Representation S-free sets for Violation Ball Max Full-Dim OPF Sets are Cones Negative Semidefinite P i Instance V2 Gap V2 Time Gap Closed Time Ex % % 2.33 Ex % % 0.59 Ex % % 0.34 Ex % % Ex % % Ex % % 0.12 Ex % 0.12 The2 2 cone Characterization for 2 2 Cuts Experiments: Setup Experiments: Results BoxQP SIAM OPT 2017 S-free sets for PolyOpt 27

40 Take-homes Thanks! SIAM OPT 2017 S-free sets for PolyOpt 28

41 Take-homes Take-homes Thanks! We introduced an oracle-based intersection cut for closed sets. Furthermore, we constructed a convergent cutting plane algorithm that uses this oracle to ping the sets. All of this is done without using any explicit structure abouts. Outer-product-free sets provide a new way to generate cuts for polynomial optimization. SIAM OPT 2017 S-free sets for PolyOpt 29

42 Thanks! Preprint available: Take-homes Thanks! SIAM OPT 2017 S-free sets for PolyOpt 30