Upload
others
View
10
Download
0
Embed Size (px)
Citation preview
Introduction Algorithm Numerical Results Conclusions
An Outer Approximation Algorithm for NonlinearMixed-Integer Programming based on Sequential
Quadratic Programming with Trust Region Stabilization
O. Exler1 T. Lehmann2 K. Schittkowski1
1University of Bayreuth - Department of Applied Computer Science
2ZIB - Konrad-Zuse-Zentrum Berlin - Department Optimization
... supported by Shell GameChanger (SIEP-EPR Rijswijk), under projectnumber 4600003917
Marseille, April 13, 2010
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 1 / 22
Introduction Algorithm Numerical Results Conclusions
Outline
1 IntroductionProblem FormulationExample: Well Relinking
2 AlgorithmMixed-Integer SQP-Trust-Region MethodLinear Outer ApproximationOuter Approximation based on MISQP
3 Numerical Results100 Academic Test Cases - Schittkowski 200955 Test Cases from Petroleum Engineering
4 Conclusions
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 2 / 22
Introduction Algorithm Numerical Results Conclusions
Problem Formulation
Mixed Integer Nonlinear Programming (MINLP):
min f (x , y) objective, e.g. costs
s.t. gj(x , y) ≥ 0 , j = 1, . . . ,m constraints, e.g. operational limits
x ∈ X , y ∈ Y
whereX := {x ∈ IRnc : xlb ≤ x ≤ xub} ,
Y := {y ∈ INni : ylb ≤ y ≤ yub} .
Assumptions:
Problem is non-convex and non-relaxable.
The model functions are continuously differentiable subject to thecontinuous variables x ∈ IRnc .
Time-consuming function evaluations.
No equality constraints for simplicity.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 3 / 22
Introduction Algorithm Numerical Results Conclusions
Problem Formulation
Mixed Integer Nonlinear Programming (MINLP):
min f (x , y) objective, e.g. costs
s.t. gj(x , y) ≥ 0 , j = 1, . . . ,m constraints, e.g. operational limits
x ∈ X , y ∈ Y
whereX := {x ∈ IRnc : xlb ≤ x ≤ xub} ,
Y := {y ∈ INni : ylb ≤ y ≤ yub} .
Assumptions:
Problem is non-convex and non-relaxable.
The model functions are continuously differentiable subject to thecontinuous variables x ∈ IRnc .
Time-consuming function evaluations.
No equality constraints for simplicity.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 3 / 22
Introduction Algorithm Numerical Results Conclusions
Example: Well Relinking
Problem: Maximize total production, three sources with differentproduction characteristics, three available sinks (nc = 3, ni = 9), simplestcase
Variables: Continuous flows x1, x2, x3 and binary decisions for splittingy1, . . . , y9
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 4 / 22
Introduction Algorithm Numerical Results Conclusions
Example: Well Relinking
Maximize: Total flowx1 + x2 + x3
Splitfactor Constraints: Define switching conditions
y1 + y2 + y3 = 1
y4 + y5 + y6 = 1
y7 + y8 + y9 = 1
Upper Bounds: For pressure at source and mass rate at sink
50y1 + 10y2 + 5y3 ≤ 90− 89x1/900 x1y1 + x2y4 + x3y7 ≤ Qmax
50y4 + 10y5 + 5y6 ≤ 80− 79x2/900 x1y2 + x2y5 + x3y8 ≤ Qmax
50y7 + 10y8 + 5y9 ≤ 70− 69x3/900 x1y3 + x2y6 + x3y9 ≤ Qmax
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 5 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
Approach
Mixed Integer Sequential Quadratic Programming (MISQP).Extension of the well-known SQP trust-region methods from continuousnonlinear optimization (Fletcher 1982, Yuan 1995).Advantages:
Continuous and integer variables are optimized directly andsimultaneously.
Approximation of L∞-penalty function.
Integer variables and computational complexity passed to QPsubproblem.
Efficient computation of a feasible solution, often close to the optimalone.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 6 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
L∞-penalty function with penalty parameter σ > 0:
Pσ(x , y) = f (x , y) + σ‖g(x , y)−‖∞ ,
Approximation:
Φk(d) := ∇f (xk , yk)Td +1
2dTBkd + σk‖(g(xk , yk) +∇g(xk , yk)Td)−‖∞
where d :=
(dc
d i
)with dc ∈ IRnc and d i ∈ INni .
Subproblem:
min Φk(d)
‖dc‖∞ ≤ ∆ck , ‖d i‖∞ ≤ ∆i
k ,
xk + dc ∈ X , yk + d i ∈ Y ,
dc ∈ IRnc , d i ∈ INni .
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 7 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
Subproblem:
min Φk(d)
‖dc‖∞ ≤ ∆ck , ‖d i‖∞ ≤ ∆i
k ,
xk + dc ∈ X , yk + d i ∈ Y ,
dc ∈ IRnc , d i ∈ INni .
Subproblem reformulation (MIQP):
min ∇f (xk , yk)Td + 12dTBkd + σkδ
δ + gj(xk , yk) +∇gj(xk , yk)Td ≥ 0 , j = 1, . . . ,m ,
‖dc‖∞ ≤ ∆ck , ‖d i‖∞ ≤ ∆i
k ,
xk + dc ∈ X , yk + d i ∈ Y , δ ≥ 0 ,
dc ∈ IRnc , d i ∈ INni , δ ∈ IR .
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 8 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
Subproblem:
min Φk(d)
‖dc‖∞ ≤ ∆ck , ‖d i‖∞ ≤ ∆i
k ,
xk + dc ∈ X , yk + d i ∈ Y ,
dc ∈ IRnc , d i ∈ INni .
Subproblem reformulation (MIQP):
min ∇f (xk , yk)Td + 12dTBkd + σkδ
δ + gj(xk , yk) +∇gj(xk , yk)Td ≥ 0 , j = 1, . . . ,m ,
‖dc‖∞ ≤ ∆ck , ‖d i‖∞ ≤ ∆i
k ,
xk + dc ∈ X , yk + d i ∈ Y , δ ≥ 0 ,
dc ∈ IRnc , d i ∈ INni , δ ∈ IR .
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 8 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
Quotient of actual and predicted improvement of Pσk(xk):
rk =Pσk
(xk , yk)− Pσk(xk + dc
k , yk + d ik)
Φk(0)− Φk(dk).
⇒ Update trust region radii dependent on rk .
Update Trust Region Parameter:
∆ck+1 =
{max(2∆c
k , 4‖dck ‖∞) , rk > 0.9
min(∆ck/4, ‖dc
k ‖∞/2) , rk < 0.1
∆ik+1 =
{max(2∆i
k , 4‖d ik‖∞) , rk > 0.9
min(∆ik/4, ‖d i
k‖∞/2) , rk < 0.1
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 9 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
General Idea
1 Compute function values f (xk , yk), g1(xk , yk),..., gm(xk , yk)
2 Approximate ∇f (xk , yk), ∇g1(xk , yk),..., ∇gm(xk , yk) by differenceformula, at grid points for integer variables
3 Update quasi-Newton Bk matrix using modified BFGS formula
4 Generate MIQP and apply trust region parameter separately forcontinuous and integer variables (∆c
k , ∆ik)
5 Solve the MIQP by any available technique, e.g., branch-and-cut
6 Add second-order correction of Yuan 1995, if necessary
Properties:
Quick computation of feasible integer solutions
Applicable to non-convex and non-relaxable programs
No global convergence, even for convex problems
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 10 / 22
Introduction Algorithm Numerical Results Conclusions
Mixed-Integer SQP-Trust-Region Method
General Idea
1 Compute function values f (xk , yk), g1(xk , yk),..., gm(xk , yk)
2 Approximate ∇f (xk , yk), ∇g1(xk , yk),..., ∇gm(xk , yk) by differenceformula, at grid points for integer variables
3 Update quasi-Newton Bk matrix using modified BFGS formula
4 Generate MIQP and apply trust region parameter separately forcontinuous and integer variables (∆c
k , ∆ik)
5 Solve the MIQP by any available technique, e.g., branch-and-cut
6 Add second-order correction of Yuan 1995, if necessary
Properties:
Quick computation of feasible integer solutions
Applicable to non-convex and non-relaxable programs
No global convergence, even for convex problems
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 10 / 22
Introduction Algorithm Numerical Results Conclusions
Linear Outer Approximation
Global stabilization by Linear Outer Approximation
Motivation:
Global lower bounds guarantee optimality forconvex MINLP problems.
Exploration of different areas of the feasibleregion.
Linear Outer Approximation:
Overestimation of feasible region.
Underestimation of objective function.
Linear relaxation of MINLP problem.
Pictures: Grossmann 2002O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 11 / 22
Introduction Algorithm Numerical Results Conclusions
Linear Outer Approximation
Master Program: Duran and Grossmann 1986, Fletcher and Leyffer 1994
min η
uk > η
η ≥ f (xp, yp) +∇f (xp, yp)T
(x − xp
y − yp
), p ∈ Tk
0 ≤ gj(xp, yp) +∇gj(xp, yp)T
(x − xp
y − yp
), j = 1, ...,m,
0 ≤ gj(xl , yl) +∇gj(xl , yl)T
(x − xl
y − yl
), j = 1, ...,m, l ∈ Sk
x ∈ X , y ∈ Y , η ∈ IR
Tk : set of previous iterates, for which MINLP for fixed yk solvable, xk isthe corresponding solution.Sk : set of previous iterates, for which MINLP not solvable, xk is thecorresponding solution of a feasibility problem.uk : current upper bound for objective function.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 12 / 22
Introduction Algorithm Numerical Results Conclusions
Outer Approximation based on MISQP
Trust Region Part
1 Solve MIQP with ∆ik = 0 fixed, giving d̂k = (d̂c
k , 0) .
2 Solve MIQP with d ik ∈ INni , giving dk = (dc
k , dik) .
3 Evaluate Pσk((xk , yk) + dk) and Pσk
((xk , yk) + d̂k) .
4 IfPσk
((xk , yk) + dk) < Pσk((xk , yk) + d̂k) ,
then use dk . (Improving mixed integer search direction)Else set dk := d̂k .
5 Calculate rk =Pσk
(xk ,yk )−Pσk((xk ,yk )+dk )
Φk (0)−Φk (dk ) .
6 Depending on rk , reduce or increase the trust region radii.
7 If Φk(0)− Φk(dk) small, convergence against stationary point, gotoStep 8 (Outer Approximation Part).Else goto Step 1.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 13 / 22
Introduction Algorithm Numerical Results Conclusions
Outer Approximation based on MISQP
Outer Approximation Part
8 Update linear outer approximation master problem by addinglinearizations and solve Master Program, giving (xOA
k , yOAk ) .
If yOAk 6= yOA
l , for all previous OA steps, let k = k + 1 and goto Step 1and repeat Trust Region Part with Step 2.
Else goto Step 1 and repeat Trust Region Part without Step 2.
Theorem
The Algorithm terminates at an optimal solution of convexMINLP-problems.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 14 / 22
Introduction Algorithm Numerical Results Conclusions
Codes
MINLP solvers:
MISQPN: Outer Approximation solver supported by MIQP search steps(preliminary implementation)
MISQPN0: MISQPN without MIQP search steps
MISQP: Mixed integer SQP trust region method
MISQPOA: External Outer Approximation solver based on MISQP
MINLPB4: Branch and bound solver based on MISQP
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 15 / 22
Introduction Algorithm Numerical Results Conclusions
Test Case Library
Collection of 100 academic MINLP test cases.Details:
nc ≤ 21,ni ≤ 100m ≤ 54, including nonlinear equality constraints.
(available:
http://www.math.uni-bayreuth.de/̃ kschittkowski/downloads.htm )
Collection of 55 problems from Petroleum Engineering.Details:
3 ≤ nc ≤ 10,9 ≤ ni ≤ 27m ≤ 30, including linear equality constraints.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 16 / 22
Introduction Algorithm Numerical Results Conclusions
100 Academic Test Cases - Schittkowski 2009
Numerical Resultscode nsucc ∆err nerr nfunc ngrad
MISQPN 74 1.19 3 714 35MISQPN0 57 5.68 19 316 21MISQP 91 0.15 1 793 29MISQPOA 96 0.05 1 6,006 102MINLPB4 91 1.16 4 111,611 4,430
Table: Test Results for a Set of 100 Academic Test Problems
nsucc - number of successful test runs, (f (xk , yk)− f ?)/|f ?| ≤ 0.01∆err - average relative deviation of computed solution from known
one, (f (xk , yk)− f ?)/|f ?| , taken over all feasible solutionsnerr - number of test runs terminated by an error messagenfunc - average number of equivalent function calls including function
calls used for gradient approximations, for successful runsngrad - average number of gradient evaluations, for successful runs
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 17 / 22
Introduction Algorithm Numerical Results Conclusions
100 Academic Test Cases - Schittkowski 2009
code nsucc nloc nerr nfunc ngrad time
MISQP/rx 90 8 2 68 18 0.02MISQP 91 8 1 793 29 15.20
Table: Performance Results of Mixed-Integer versus Continuously Relaxed TestProblems
Observation: At least 10 % of our test problems are non-convex.
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 18 / 22
Introduction Algorithm Numerical Results Conclusions
55 Test Cases from Petroleum Engineering
Numerical Results
code nsucc ∆err nerr nfunc ngrad
MISQPN 47 0.05 8 14,064 648MISQPN0 33 0.26 21 17,340 807MISQP 34 0.08 1 1,797 82MISQPOA 50 0.03 0 30,359 1,925MINLPB4 55 0.0 0 176,460 8,107
Table: Performance Results for a Set of 55 Well Relinking and Gas Lift TestProblems
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 19 / 22
Introduction Algorithm Numerical Results Conclusions
Conclusions I
MISQPN - efficient if binary variables are present (preliminaryversion), MIQP search steps increase robustness (see MISQPN0)
MISQP - extremely efficient computation of feasible solutions, lessrobust especially for binary problems
MISQPOA - MISQP called inside, less efficient, but more robust byouter approximations, especially for binary variables.
MINLPB4 - very robust, but highly inefficient, not applicable withtime-consuming function evaluations
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 20 / 22
Introduction Algorithm Numerical Results Conclusions
Conclusions II
Development of efficient mixed-integer nonlinear optimizationalgorithms
Applicable for solving non-convex and non-relaxable programs
Efficient computation of feasible solutions
Large internal calculation times (MIQP solutions), acceptable due toexpensive simulations in real world applications
Applicable without analytical derivatives
Codes in practical use (Shell, Pace Aerospace Engineering, GeneralElectric, ...)
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 21 / 22
Introduction Algorithm Numerical Results Conclusions
The End
Thank you for your attention!
O. Exler (University of Bayreuth) Outer Approximation based on MISQP EWMINLP April 12-16, 2010 22 / 22