Convex Optimization : Conic Versus Functional Form

Size: px

Start display at page:

Download "Convex Optimization : Conic Versus Functional Form"

Abigayle Fitzgerald
5 years ago
Views:

1 Convex Optimization : Conic Versus Functional Form Erling D. Andersen MOSEK ApS, Fruebjergvej 3, Box 16, DK 2100 Copenhagen, Blog: Linkedin: e.d.andersen@mosek.com WWW:

2 2 / 33

3 Topic Topic Considerations about the forms Conic optimization. A recap Basic cone types Comments Given a convex optimization problem then is the classical form min c(x) st a(x) 0 or the conic form min c T x st Ax = b, x K. preferable from an user perspective? Assumptions c(x) is convex anda(x) is concave. K is a convex cone. We can only solve convex problems efficiently. 3 / 33

4 Considerations about the forms Topic Considerations about the forms Conic optimization. A recap Basic cone types Comments Robustness. Accuracy of the solution. How sensitive is the optimizer to the problem data. Solution time. Does the form help the user to build a good. Does it prevent nonconvex problems because checking convexity is hard. Ease of use. 4 / 33

5 Conic optimization. A recap Topic Considerations about the forms Conic optimization. A recap Basic cone types Comments Primal problem: (CO P ) min c T x st Ax = b, x K 1 K 2 K k, where eachk k is a cone (closed, pointed and solid). Dual problem: (CO P ) max b T y st A T y +s = c, s K1 K 2 K k, where eachk k is the dual convex cone i.e. K k {s : x T s 0, x K}. 5 / 33

6 Basic cone types Topic Considerations about the forms Conic optimization. A recap Basic cone types Comments Linear: Quadratic: K q := K l := {x R : x 0} x Rn : x 1 n j=2 x 2 j 6 / 33

7 Comments Topic Considerations about the forms Conic optimization. A recap Basic cone types Comments We will only deal with the above 2 self-dual cones. A quote from Stephen Boyd of Stanford: Almost all convex optimization problem can be formulated using a combination of linear, quadratic, semi-definite and the exponential cones. See also the book of Ben-Tal and Nemirovski for conic quadratic representable sets. 7 / 33

8 8 / 33

9 Convex quadratic optimization Convex quadratic optimization continued Practical issues Functional formulation: min 1 2 xt Qx+c T x st Ax = b, x 0. Given convexity then there exists a H such that Q = HH T. Conic reformulation min 1 2 t+ct x st Ax = b, H T x y = 0, s = 1, y 2 2st, x 0. 9 / 33

10 continued Convex quadratic optimization continued Practical issues Conic reformulation requires more constraints and variables. The addition is highly structured and sparse. Conic form is convex by construction. Q is frequently negative semi-definite due to rounding errors or mistakes. How to deal with: [ ǫ ]? In practiceqis frequently low rank and dense. Implying the conic reformulation requires much less space to store. 10 / 33

11 Practical issues Convex quadratic optimization continued Practical issues Many users of the nonlinear capabilities in MOSEK: Do believe their nonlinear problems are convex. Do believe convexity is not important. Almost always make mistakes. (Errors in function or gradient computations). Checking and verifying convexity is an issue even for. Checking convexity and smoothness for arbitrary black box functions are hard. Data for the conic form is very simple. Just vectors, matrices and list of indexes. 11 / 33

12 12 / 33

13 The conic s The available s determines which form is preferable. Only primal-dual interior-point s are considered. They are efficient in practice. Polynomial complexity holds: For some. All conic. Claim: The conic form is more efficient and robust than functional form. Justification of claim: Follows. 13 / 33

14 functional form The conic s Problem: Assumptions: 1. x R n min c(x) st a i (x) 0, i, x c is convex and twice differentiable. 3. a i is concave and twice differentiable. 14 / 33

15 The conic s Lagrange function wherey 0. The dual problem: s R n. Alternatively: L(x,y) c(x) y T a(x), max L(x,y) x T s st (x) L(x,y) T = s, x,y,s 0, max L(x,y) x L(x,y)x st x L(x,y) T = s, x,y,s / 33

16 The conic s (MCP) min (x;y) T (s;z) st a(x) = z, st x L(x,y) T = s, x,z,y,s 0. wherez R m is slack variables. Note: The KKT is a monotone complementarity problem. Primal-dual interior-point type. Difficulties with detecting infeasibility! Feasibility constraints are potentially highly nonlinear! 16 / 33

17 The conic s denoted (HMCP): min (x;τ) T (s;κ) st τ a(x/τ) = z, primal feasibility τ x L(x/τ,y/τ) T = s, dual feasibility, x T x L(x/τ,y/τ) T y T a(x/τ) = κ, zero gap, x,z,τ,s,y,κ 0. Has two additional variablesτ,κ and one additional constraint. (HMCP) is a homogeneous monotone complementarity problem. Always has solution (0). 17 / 33

18 The conic s Theorem: (Andersen and Ye 95) Let (x,z,y,τ,s,κ ) be a maximal complementarity solution to (HMCP). Then (MCP) has a solution if and only if τ > 0. In this case, (x,z,y,s )/τ is a complementarity solution for(mcp). Conclusion: Compute a maximal solution to (HMCP). 18 / 33

19 The conic s Algorithm: Apply a primal-dual interior-point method to (HMCP) i.e. apply Newtons method to the perturbed optimality : τ x L(x/τ,y/τ) T = s, τa(x/τ) = z, x T x L(x/τ,y/τ) T y T a(x/τ) = κ, Xs = µe, Zy = µe, τκ = µ, x,z,τ,s,y,κ > 0. µ > 0: Is a parameter. Primal feasibility cond. is nonlinear. Dual feasibility cond. is nonlinear. Perturbed complementarity cond. is nonlinear. 19 / 33

20 The Step 0: Choose (x;z;τ;y;s;κ) > 0. Step 1: Choose parameters. Step 2: Compute the residuals: The where r r D r G r P s τg κ+x T g y T a(x/τ) z τa(x/τ) w (x;z;τ;y;s;κ) J a(x/τ), g x L(x/τ,y/τ) T H 2 xl(x/τ,y/τ)., conic s 20 / 33

21 The conic s Step 3: Solve the Newton equation system: and where H v2 J T (v 1 ) T H G (v 3 ) T J v 3 0 v 1 v 2 v 3 H g d x d τ d y d s d κ d z Sd x +Xd s = Xs+γµe, Yd z +Zd y = Zy +γµe, κd τ +τd κ = τκ+γµ, c(x/τ) T Hx/τ, c(x/τ) T Hx/τ, a(x/τ) Jx/τ, x/τ T H(x/τ), µ (x;z;τ) T (s;y;κ)/(n+m+1). = ηr 21 / 33

22 The conic s η = (γ 1) whereγ [0,1] is an ic parameters. Step 4: Update. Update of slacks: or where (x + ;τ + ;y + ;) = (x;τ;y)+α(d x ;d τ ;d y ) (z + ;s + ;κ + ) = (z +αd z ;s+αd s ;κ+αd κ ) z + = ((1+αη)r P +τ + a(x + /τ + ), s + = ((1+αη)r D +g + ), κ + = ((1+αη)r G +g + κ, g + τ + x L(x + /τ +,y + /τ + ) T g κ + (x + ) T x L(x + /τ +,y + /τ + ) T (y + ) T a(x + /τ + ) 22 / 33

23 The conic s Step 5: Repeat step 2 to 4. Issues: Gap and residuals are reduced by the factor(1+αη) in every iteration (given the nonlinear ). Polynomial complexity holds under certain assumptions. OBSERVE: z + = ((1+αη)r P +τ + a(x + /τ + ), s + = ((1+αη)r D +g + ). κ + = ((1+αη)r G +g + κ). Does not always work well!!! 23 / 33

24 The conic s Alternative. A linear z + = x+αd z, s + = s+αd s κ + = κ+αd κ but then a merit function is required e.g. two norm of the residuals. Works most of the time. Can lead to very slow convergence and robustness problems. A major weakness. Nonlinear primal and dual feasibility constraints are the cause of the problems. 24 / 33

25 conic quadratic problems The conic s : (H) Ax bτ = 0, (primal feasibility) A T y +s cτ = 0, (dual feasibility) c T x+b T y κ = 0, (zero gap) (x;τ) K,(s;κ) K In linear case suggested suggested Goldman and Tucker. Later rediscovered by Mizuno, Todd and Ye. All solutions satisfies: x T s+τκ = 0. If and only ifτ > 0, then an optimal solution exists. Find a solution such that τ +κ > 0. Linear primal and dual feasibility!!! Algorithmic framework: The Nesterov-Todd direction. 25 / 33

26 The conic s NT : Ad x bd τ A T d y +d s cd τ c T d x +b T d y d κ XT(ΘW) 1 d s + STΘWd x τd κ +κd τ whereη := γ 1 andγ [0,1). New iterate: x + x τ + τ y + = y s + s κ + κ = η(ax bτ), = η(a T y +s cτ), = η( c T x+b T y κ), = X Se+γµe, = τκ+γµ. +α d x d τ d y d s d κ. 26 / 33

27 The conic s Fact: Ax + bτ + = (1+αη)(Ax bτ), A T y + +s + cτ + = (1+αη)(A T y +s cτ), c T x + +b T y + κ + = (1+αη)( c T x+b T y κ), d T xd T s +d τ d κ = 0, (x + ) T s + +τ + κ + = (1+αη)((x) T s+τκ). Observations: NT can be seen as Newtons method. The method is symmetric! The complementarity gap is reduced by a factor of (1+αη) [0,1). The infeasibility is reduced by the same factor. No merit function is required. 27 / 33

28 s The conic s Both s are called primal-dual interior-point. Only in the conic form are the feasibility constraints linear. The functional one is more like primal the KKT system The conic is truly symmetric i.e. you can flip the primal and dual problem. Due to issues with the merit function then the functional form is not as robust as conic one. 28 / 33

29 The conic s Conic form: Functional form: min st min st y x = y x y, x R y x = y x 2 y y, y 0, x,y R 29 / 33

30 MOSEK v6 on the conic formulation The conic s ITE PFEAS DFEAS KAP/TAU POBJ DOBJ MU 0 1.0e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e / 33

31 MOSEK v6 on the ulation The conic s ITE PFEAS DFEAS KAP/TAU POBJ DOBJ MU 0 1.0e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e-012 Performance on the ulation is bad. Merit function forces the alg. to take small steps. Behavior seen occasionally on large instances. Other interior-point optimizers may do better of course. 31 / 33

32 Summary The conic s The (restricted) conic formulation of nonlinear problems is advantageous because The problem data is simpler to deal in a software sense and no blackbox functions are required. It is impossible to formulate nonconvex problems. The available is much better. It easy to dualize the conic formulation. Caveats: The conic formulation is usually bigger in terms of number of variables and constraints but also highly structured. Usually the storage and time spend per iteration will be same or less for the conic formulation. 32 / 33

33 Conclusion The conic form is better than the! The conic s 33 / 33

On implementing a primal-dual interior-point method for conic quadratic optimization

On implementing a primal-dual interior-point method for conic quadratic optimization E. D. Andersen, C. Roos, and T. Terlaky December 18, 2000 Abstract Conic quadratic optimization is the problem of minimizing