Optimization Problems with Equilibrium Constraints · Primal-dual MPEC system with x 1;x 2 in...
Transcript of Optimization Problems with Equilibrium Constraints · Primal-dual MPEC system with x 1;x 2 in...
Optimization Problems with EquilibriumConstraints
GIAN Short Course on Optimization:Applications, Algorithms, and Computation
Sven Leyffer
Argonne National Laboratory
September 12-24, 2016
Outline
1 Solving MPECs as NLPs
2 Convergence for Sequential Quadratic Programming Methods
3 Convergence for Interior-Point Methods
4 An SLPEC-EQP ApproachCounter Example for SQPECSLPEC MethodAccelerating Local Convergence
2 / 32
Solving MPECs as NLPs
Mathematical Program with Equilibrium Constraints (MPEC)minimize
x ,yf (x , y)
subject to c(x , y) ≥ 00 ≤ y ⊥ F (x , y) ≥ 0
Equivalent smooth (lazy) nonlinear program (NLP):minimize
x ,yf (x , y)
subject to c(x , y) ≥ 0F (x , y) = s, s ≥ 0, y ≥ 0 and yT s ≤ 0
3 / 32
Switching Notation
To understand convergence analysis, we switch notation:x = (x0, x1, x2):
minimizex
f (x)
subject to c(x) ≥ 00 ≤ x1 ⊥ x2 ≥ 0
Equivalent smooth nonlinear program (NLP):minimize
xf (x)
subject to c(x) ≥ 0x1 ≥ 0, x2 ≥ 0, xT
1 x2 ≤ 0
Now examine convergence properties of NLP solvers ...
4 / 32
A Nonlinear Programming Approach
Replace equilibrium 0 ≤ x1 ⊥ x2 ≥ 0 by X1x2 ≤ 0 or xT1 x2 ≤ 0
⇒ standard nonlinear program (NLP)
(NLP)
minimize
xf (x)
subject to c(x) ≥ 0x1, x2 ≥ 0
X1x2 ≤ 0x
x
1
2
Advantage: standard (?) NLP; use large-scale solvers ...Snag: nonlinear program (NLP) violates standard assumptions!
5 / 32
Strong Stationarity & Unbounded MultipliersExample x∗ = (0, 1):{
minx
12(x1 − 1)2 + (x2 − 1)2
s.t. x1, x2 ≥ 0, x1x2 ≤ 0
first order conditions:(−1
0
)=
(ν10
)−(ξ0
)ν1 multiplier of x1 ≥ 0; ξ multiplier of x1x2 ≤ 0.
Equivalent NLP (x1x2 ≤ 0) violates MFCQ ⇒ unboundedmultipliers
x
x
1
2
x*
ν
ξ
−1= −ν ξ
multipliers form a ray ⇒ ∃ bounded multipliers
6 / 32
Outline
1 Solving MPECs as NLPs
2 Convergence for Sequential Quadratic Programming Methods
3 Convergence for Interior-Point Methods
4 An SLPEC-EQP ApproachCounter Example for SQPECSLPEC MethodAccelerating Local Convergence
7 / 32
The Relaxed NLP
Define index sets
X1 := {i : x∗1i = 0} & X2 := {i : x∗2i = 0} ,
complements X⊥j := {1, . . . , p} − Xj
⇒ relaxed NLP given by
minimizex
f (x)
subject to c(x) ≥ 0x1j = 0 ∀j ∈ X⊥2x2j = 0 ∀j ∈ X⊥1x1, x2 ≥ 0
... i.e. µi multiplier of “equality” constraints
8 / 32
Equivalence to KKT Conditions
KKT conditions of equivalent NLP: ∃λ∗, ν∗1 , ν∗2 , ξ∗ ≥ 0
∇f (x∗)−∇c(x∗)Tλ∗ −
0ν∗1 − X ∗2 ξ
∗
ν∗2 − X ∗1 ξ∗
= 0 1st order
c(x∗) ≥ 0, x∗1 ≥ 0, x∗2 ≥ 0 and X ∗1 x∗2 ≤ 0 primal feas.
c(x∗)Tλ = x∗T
1 ν∗1 = x∗T
2 ν∗2 = 0 compl. slack.
... ξ > 0 allows µ1 < 0
... multipliers of relaxed NLP µ1 = ν1 − X ∗2 ξ, and µ2 = ν2 − X ∗1 ξ⇒ KKT multipliers bounded if ‖ξ∗‖ <∞
9 / 32
Convergence of SQP for MPECs
Sequential Quadratic Programming (SQP) ... compute step d
minx
f (x)
s.t. c(x) ≥ 0x1 ≥ 0x2 ≥ 0X1x2 ≤ 0
→
mind∇f T
k d + 12dTHkd
s.t. ck +∇cTk d ≥ 0
xk1 + d1 ≥ 0xk2 + d2 ≥ 0Xk1xk2 + Xk1d2 + Xk2d1 ≤ 0
where Hk ' ∇2fk −∑λi∇2ck Hessian of the Lagrangian.
Set xk+1 = xk + d & update multiplier estimates
Two cases: ∃k : Xk1xk2 = 0 ... or ... Xk1xk2 > 0, ∀k
10 / 32
Convergence of SQP for MPECs
Sequential Quadratic Programming (SQP) ... compute step d
minx
f (x)
s.t. c(x) ≥ 0x1 ≥ 0x2 ≥ 0X1x2 ≤ 0
→
mind∇f T
k d + 12dTHkd
s.t. ck +∇cTk d ≥ 0
xk1 + d1 ≥ 0xk2 + d2 ≥ 0Xk1xk2 + Xk1d2 + Xk2d1 ≤ 0
where Hk ' ∇2fk −∑λi∇2ck Hessian of the Lagrangian.
Set xk+1 = xk + d & update multiplier estimates
Two cases: ∃k : Xk1xk2 = 0 ... or ... Xk1xk2 > 0, ∀k
10 / 32
Convergence of SQP Part 1: Xk1xk2 = 0
wlog have xk1 = 0 (and for simplicity assume xk2 > 0)
⇒ QP contains constraints
xk1 + d1 ≥ 0xk2 + d2 ≥ 0
Xk1xk2 + Xk2d1 + Xk1d2 ≤ 0
⇒d1 ≥ 0
xk2 + d2 ≥ 0Xk2d1 ≤ 0
⇒ d1 = 0
⇒ xk+11 = xk1 + d1 = 0 ... stay on same axis
⇒ same tangent cone as NLP with x1 = 0 ... relaxed NLP⇒ fast local convergence
11 / 32
Convergence of SQP Part 2: Xk1xk2 > 0
wlog x∗1 = 0, but Xk1xk2 > 0, i.e. off axis
QP picks nonsingular basis, subset of 0 0∇ck I Xk2
0 Xk1
Assume all QPs consistent ... 2 cases:
�� �� �������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
�����������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
��������������������������������
��������������������������������
case 1: true subset ⇒ non-singular ⇒ quadratic convergence
case 2: full set ⇒ xk1 > 0 (otherwise singular)⇒ X k+1
1 xk+12 = 0 now see Part (1) as before ...
12 / 32
Convergence of SQP Part 2: Xk1xk2 > 0
wlog x∗1 = 0, but Xk1xk2 > 0, i.e. off axis
QP picks nonsingular basis, subset of 0 0∇ck I Xk2
0 Xk1
Assume all QPs consistent ... 2 cases:
�� �� �������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
�����������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
��������������������������������
��������������������������������
case 1: true subset ⇒ non-singular ⇒ quadratic convergence
case 2: full set ⇒ xk1 > 0 (otherwise singular)⇒ X k+1
1 xk+12 = 0 now see Part (1) as before ...
12 / 32
Consistency of QP Approximations
Are QPs always consistent for MPECs?
NO! Linearization can be inconsistent arbitrarily close to solution
minimizex
x1 + x2
subject to x22 ≥ 1
x1 ≥ 0x2 ≥ 0x1 x2 ≤ 0
x (k)
1x
2x
x*
generic problem ⇒ solvers take arbitrary steps
13 / 32
Consistency of QP Approximations
Relax linearization of X1x2 ≤ 0 ...... heuristic for infeasible QPs (0 < δ, κ < 1 constants)
Xk1xk2 + Xk2d1 + Xk1d2 ≤ δ(
xTk1xk2
)1+κe
x (k)
1x
2x
x*
... works in well practice with δ = 0.1, κ = 1
14 / 32
The Slacks Matter!!!
How important was the introduction of slack variables?
Consider MPEC without slacks ...
(P)
minimize
z−x1 − 1
2x2
subject to x1 + x2 ≤ 20 ≤ x2
1 − x1 ⊥ x2 ≥ 0 .
with solutions (2, 0)T with f ∗ = −2 and (0, 2)T with f ∗ = −1
1
1 2
2
z1
z2
f−
z(0)
zoo
Start (−ε, t)T
Nonstationary limit(0, t)T for any t.
Avoid failure with slacks
15 / 32
Outline
1 Solving MPECs as NLPs
2 Convergence for Sequential Quadratic Programming Methods
3 Convergence for Interior-Point Methods
4 An SLPEC-EQP ApproachCounter Example for SQPECSLPEC MethodAccelerating Local Convergence
16 / 32
Interior Point Penalty Methods for MPECs
Equivalent NLP: minimize
xf (x)
subject to c(x) ≥ 0x1 ≥ 0, x2 ≥ 0,xT1 x2 ≤ 0
Consider `1 penalty of complementarity constraintminimize
xf (x)+πxT
1 x2
subject to c(x) ≥ 0x1 ≥ 0, x2 ≥ 0
... form primal-dual system with a twist ...
17 / 32
Interior Point Penalty Methods for MPECs
Primal-dual MPEC system with x1, x2 in primal form∇f (x)−∇c(x)Tλ−
0
µX−11 e − X2π
µX−12 e − X1π
= 0
c(x)− s = 0Sλ = µe
Algorithm I: Interior Penalty Method for MPECs
1 Choose barrier parameter µk , and tolerance εk2 Solve PD system to tolerance εk and ensure
‖min{xk1, xk2}‖ ≤√εk by adjusting πk
18 / 32
Interior Point Penalty Methods for MPECs
Theorem
If Algorithm I generates an infinite sequence, then:
1 xk → x∗ is feasible,
2 LICQ for relaxed NLP ⇒ x∗ is C-stationary,
3 πkxki → 0 ⇒ x∗ strongly stationary,
4 superlinear convergence for suitable barrier updates
Practical implementation
dynamic penalty πk update during inner iteration
non-monotone reduction of complementarity: πj = 10πj if,
x jT
1 x j2 > 0.9 max
{x(j−1)T1 x
(j−1)2 , . . . , x
(j−m+1)T
1 x(j−m+1)2
}avoid trouble with badly scaled MPECs
19 / 32
Relaxed Interior Point Methods for MPECs
Perturb rhs of complementarity constraint ... X1x2 ≤ Cµe
... where µ > 0 barrier parameter ⇒ primal dual system ...
∇f (x)−∇c(x)Tλ−
0
µX−11 e − X2ξ
µX−12 e − X1ξ
= 0
c(x)− s = 0Sλ = µe
X1x2 + t = CµeT ξ = µe
⇒ central path (x(µ), ν(µ), ξ(µ)) for µ > 0[Raghunathan and Biegler, 2002, Liu and Sun, 2002]
20 / 32
Relaxed Interior Point Methods for MPECs
Compare relaxation and penalization
ξ = πt = Cµe − X1x2
T ξ = µe
⇒ Penalization ⇔ Relaxation, if
πi =µ
µCi − x1ix2ior Ci =
µ+ πix1ix2iµπi
... convergence proofs carry over!
21 / 32
Interior Point Method with Two Sided Relaxation
Clever idea by Friedlander, de Miguel & Scholtes [2003]:
MPECs have no strict interior
Relax X1x2 ≤ τe ⇒ interior → 0⇒ relax X1x2 ≤ τe
and x1 ≥ −δe, x2 ≥ −δe
Adjust τ , δ as µ→ 0
Theorem
In limit τ → 0 or δ → 0 but not both⇒ relaxed problem has non-empty interior in limit⇒ interior point methods faster & more robust
MPEC multiplier µi < 0 ⇒ reduce τi ↘ 0 ...
22 / 32
Outline
1 Solving MPECs as NLPs
2 Convergence for Sequential Quadratic Programming Methods
3 Convergence for Interior-Point Methods
4 An SLPEC-EQP ApproachCounter Example for SQPECSLPEC MethodAccelerating Local Convergence
23 / 32
SQPEC Approach [Scholtes, 2004]
Sequential QPEC approach (similar to piecewise SQP)
minimized
g (k)T d + 12dTH(k)d
subject to c(k) + A(k)T d ≥ 0,
0 ≤ x(k)1 + d1 ⊥ x
(k)2 d2 ≥ 0
where g (k) = ∇f (x (k)) and A(k) = ∇c(x (k), y (k)),
Solve sequence of QPECs, set x (k+1) = x (k) + d
Theorem [Scholtes, 04]: Local B-stationary convergence.
SQPEC has correct tangent cone ⇒ global convergence???
24 / 32
No! Counter Example for SQPECConsiderminimize (x1 − 1)2 + x3
2 + x22 subject to 0 ≤ x1 ⊥ x2 ≥ 0
SQPEC: x (k+1) =(
0, 3x(k)2
2 /(6x(k)2 + 2)
)→ (0, 0) spurious
25 / 32
A Sequential LPEC Methodwhile (not optimal) begin
1 Compute step d from LPEC subproblem
minimized
g (k)T d
subject to c(k) + A(k)T d ≥ 0,
0 ≤ x(k)1 + d1 ⊥ x
(k)2 + d2 ≥ 0
‖d‖∞ ≤ ∆k trust-region
2 if x (k) + d acceptable thenx (k+1) = x (k) + d & increase TR ∆(k+1) = 2 ∗∆k
else x (k+1) = x (k) & decrease TR ∆(k+1) = ∆k/2
end
1 Like steepest descend: Can we speed up convergence?
2 When is x (k) + d acceptable?
3 How do we solve the LPEC subproblem?
26 / 32
Spurious A/M-Stationarity Revisited
Consider min (x1 − 1)2 + x32 + x2
2 subject to 0 ≤ x2 ⊥ x1 ≥ 0
SLPEC pivots through (0, 0) ... get onto x1-axis
SLPEC converges to B-stationary limit (1, 0)
... cannot get stuck in spurious stationary points
27 / 32
Spurious A/M-Stationarity Revisited
Consider min (x1 − 1)2 + x32 + x2
2 subject to 0 ≤ x2 ⊥ x1 ≥ 0
����������
����������
��������
��������
��������
��������
��������
��������
trust−region
SLPEC pivots through (0, 0) ... get onto x1-axis
SLPEC converges to B-stationary limit (1, 0)
... cannot get stuck in spurious stationary points
27 / 32
Spurious A/M-Stationarity Revisited
Consider min (x1 − 1)2 + x32 + x2
2 subject to 0 ≤ x2 ⊥ x1 ≥ 0
����������
����������
��������
��������
��������
��������
��������
��������
���������������
���������������
trust−region
SLPEC pivots through (0, 0) ... get onto x1-axis
SLPEC converges to B-stationary limit (1, 0)
... cannot get stuck in spurious stationary points
27 / 32
Spurious A/M-Stationarity Revisited
Consider min (x1 − 1)2 + x32 + x2
2 subject to 0 ≤ x2 ⊥ x1 ≥ 0
����������
����������
��������
��������
��������
��������
��������
��������
���������������
���������������
trust−region
SLPEC pivots through (0, 0) ... get onto x1-axis
SLPEC converges to B-stationary limit (1, 0)
... cannot get stuck in spurious stationary points
27 / 32
Accelerating Local Convergence
28 / 32
Equality Constrained Quadratic Program (EQP)
Given active set estimate from LPEC step d :
Ac(d) :={
i : c(k)i + a
(k)T
i d = 0}
A1(d) :={
i : x(k)1i + d1i = 0
}A2(d) :=
{i : x
(k)2i + d2i = 0
}solve corresponding equality QP
EQPk(d)
minimize
sg (k)T s + 1
2sTH(k)s
subject to c(k)i + a
(k)T
i s = 0, ∀i ∈ Ac(d)
x(k)1i + s1i = 0, ∀i ∈ A1(d)
x(k)2i + s2i = 0, ∀i ∈ A2(d)
for 2nd order step s.
29 / 32
A Filter Method for MPECs
MPEC have three competing aims
1 Minimize f (x , y)
2 Minimize h(x , y) := ‖c−(x , y)‖ ... more important
3 Minimize hc(x , y) := ‖min(x1, x2)‖ ... most important
... for plots, let h(x) := h(x , y) + hc(x , y)
c(x)
f(x)
dom
inate
d
h(x) =
(h , f )kk
Borrow concept of domination frommulti-objective optimization
(hk , hck , fk) dominates (hl , h
cl , fl)
iff hk ≤ hl & hck ≤ hc
l & fk ≤ fl
i.e. (x (k), y (k)) at least as good as(x (l), y (l))
30 / 32
Global Convergence to B-Stationarity
Assumptions:
MPEC-MFCQ (i.e. every piece satisfies MFCQ) weak
x (k) remain in compact set strong
f , c twice continuously differentiable
Theorem
Outcome of SLPEC is one of:
1 restoration phase fails to find feasible point, or
2 d = 0 solves LPEC ⇒ B-stationary, or
3 limit is B-stationary.
Proof: exploit fact that LPEC ≡ disjunctive LPs
31 / 32
Conclusions
Considered convergence of four classes of methods for MPECs1 SLPEC-EQP Method
Method of choice, but LPEC hard to solveDeveloping active-set type solver for LPECs⇒ base on standard LP solvers ...
2 Sequential Quadratic Programming Methods
Often works very well ... my preferred methodFails to converge or converges slowly for degenerate MPECs
3 Interior-Point Methods
Works mostly well ... not as robust as SQPFails to converge or converges slowly for degenerate MPECs
4 Sequential penalization or regularization methods
Not as effective as SQP or IPM above... solve sequence of NLPs versus a single one!Fails to converge or converges slowly for degenerate MPECs
32 / 32
Anitescu, M. (2000).On solving mathematical programs with complementarityconstraints as nonlinear programs.Preprint ANL/MCS-P864-1200, MCS Division, ArgonneNational Laboratory, Argonne, IL, USA.
Anjos, M. and Vanelli, A. (2002).A new mathematical programming framework for facilitylayout problems.Technical Report UW-E&CE#2002-04,, Department ofElectrical & Computer Engineering, University ofWaterloo,, Canada,.
Bard, J. (1988).Convex two-level optimization.Mathematical Programming, 40(1):15–27.
Benson, H., Shanno, D. F. and Vanderbei, R. V. D. (2003).LOQO: An Interior-Point Methods for NonconvexNonlinear Programming.
32 / 32
Talk at ISMP-2003.
DeMiguel, V., Friedlander, M.P., Nogales, F.J. andScholtes, S. (2003).A superlinearly convergent interior point method forMPECs.tALK AT ISMP-2003.
Ferris, M., Meeraus, A., and Rutherford, T. (1999).Computing Wardropian equilibrium in a complementarityframework.Optimization Methods and Software, 10:669–685.
Ferris, M. and Tin-Loi, F. (1999a).Limit analysis of frictional block assemblies as amathematical program with complementarity constraints.Mathematical Programming Technical Report 99-01,University of Wisconsin.
Ferris, M. and Tin-Loi, F. (1999b).
32 / 32
On the solution of a minimum weight elastoplastic probleminvolving displacement and complementarity constraints.Computer Methods in Applied Mechanics and Engineering,174:107–120.
Ferris, M. and Pang, J. (1997).Engineering and economic applications of complementarityproblems.SIAM Review, 39(4):669–713.
Fletcher, R., Leyffer, S., Ralph, D., and Scholtes, S. (2002).Local convergence of SQP methods for mathematicalprograms with equilibrium constraints.Numerical Analysis Report NA/209, Department ofMathematics, University of Dundee, Dundee, DD1 4HN,UK.
Hearn, D. and Ramana, M. (1997(?)).Solving congestion toll pricing models.
32 / 32
Technical report, Department of Industrial and SystemsEngineering, University of Florida,http://www.ise.ufl.edu/hearn/crt.ps.
Hobbs, B., Metzler, C., and Pang, J.-S. (2000).Strategic gaming analysis for electric power systems: Anmpec approach.IEEE Transactions on Power Systems, 15(2):638–645.
Huang, J. and Pang, J.-S. (1999).A mathematical programming with equilibrium approach tothe implied volatility surface of American options.Technical report, Dept. of Mathematical Sciences, TheJohns Hopkins University, Baltimore, Maryland 21218-2682,USA.
Liu, X. and Sun, J. (2002).Generalized stationary points and an interior point methodfor mathematical programs with equilibrium constraints.Preprint, Singapore, School of Business.
32 / 32
Luo, Z.-Q., Pang, J.-S., and Ralph, D. (1996).Mathematical Programs with Equilibrium Constraints.Cambridge University Press.
Outrata, J., Kocvara, M., and Zowe, J. (1998).Nonsmooth Approach to Optimization Problems withEquilibrium Constraints.Kluwer Academic Publishers, Dordrecht.
J.-S. Pang and M. Fukushima (2002).Quasi-Variational Inequalities, Generalized Nash Equilibria,and Multi-Leader-Follower Games.Preprint The Johns Hopkins Universit, Baltimore, USA.
Pieper, H. (2001).Algorithms for Mathematical Programs with EquilibriumConstraints with Applications to deregulated electricitymarkets.PhD thesis, Department of Management Science andEngineering, Stanford University.
32 / 32
Raghunathan, A. and Biegler, L. T. (2002).MPEC formulations and algorithms in process engineering.Technical report, CMU Chemical Engineering.
Rico-Ramirez, V. and Westerberg, A. (1999).Conditional modeling. 2. solving using complementarity andboundary-crossing formulations.Industrial Engineering and Chemistry Research, 38:531–553.
Scholtes, S. (2001).Convergence properties of regularization schemes formathematical programs with complementarity constraints.SIAM Journal on Optimization, 11(4):918–936.
Stackelberg, H. V. (1952).The Theory of Market Economy.Oxford University Press.
Tin-Loi, F. and Que, N. (2002).Nonlinear programming approaches for an inverse problemin quasibrittle fracture.
32 / 32
Mechanical Sciences, 44:843–858.
Wilmott, P., Dewyne, J., and Howison, S. (1993).Option Pricing: Mathematical Models and Computation.Oxford Financial Press.
32 / 32