Penalty Alternating Direction Methods for Mixed- Integer Optimization: A New View on Feasibility Pumps

Transcription

1 Penalty Alternating Direction Methods for Mixed- Integer Optimization: A New View on Feasibility Pumps Björn Geißler, Antonio Morsi, Lars Schewe, Martin Schmidt FAU Erlangen-Nürnberg, Discrete Optimization Workshop on Combinatorial Optimization, Aussois,

2 Feasibility Pumps for MIPs Fischetti, Glover, Lodi (2005), Bertacco, Fischetti, Lodi (2007) min x s.t. c x Ax b x i Z for all i I Create two sequences (x k ) k satisfies the inequalities, (y k ) k is integral for all i I. Minimize distance of pairs x k and y k If you get stuck, perturb (at random). x k = y k : you are done. Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

3 Goal Understand feasibility pump algorithms Does the procedure converge? Can we characterize the points to which it converges? Specifically... What is the role of the perturbation step? Why randomize? In this talk... A feasibility pump variant with no randomization for MIP and MINLP. Characterization of the points to which the method converges Observation: The standard feasibility pump only has short cycles (see also Dey et al. 2016) Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

4 The Basic Feasibility Pump for MIPs Compute x 0 argmin{c x : x P}. if x 0 is integer feasible then return x 0 y 0 x 0 and k 0. while not termination condition do Compute x k+1 argmin{ x I y k I 1 : x P}. if x k+1 is integer feasible then return x k+1 y k+1 x k+1. if algorithm stalls or cycles then perturb y k+1 k k + 1 Stalling : x k = x k+1, y k = y k+1. Cycling : x k = x k+l, y k = y k+l for l > 1. Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

5 The Idealized Feasibility Pump for MIPs Compute x 0 argmin{c x : x P}. if x 0 is integer feasible then return x 0 y 0 x 0 and k 0. while True do Compute x k+1 argmin{ x I y k I 1 : x P}. y k+1 x k+1. k k + 1 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

6 How to analyze feasibility pumps? Treat idealized FP as special case of other methods Frank-Wolfe method (De Santis et al. 2013, 2014, Eckstein and Nediak 2007) Proximal point method (Boland et al. 2012) Successive projection method (D Ambrosio et al. 2012) Change randomization step of basic FP Dey, Iroume, Molinaro, and Salvagnin (2016) This talk First approach: Interpret idealized FP as Alternating Direction Method Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

7 Mixed-Integer Nonlinear Problems min x f (x) s.t. h(x) 0 x i Z [l i, u i ] for all i I Convex MINLPs Bonami, Goncalves (2012): direct extension of feasibility pump for MIPs to convex MINLPs Bonami et al. (2009): rounding step replaced by MIP relaxation (OA) of the convex MINLP; high computational effort but inheritance of OA theory Nonconvex MINLPs D Ambrosio et al. (2010): first feasibility pump for nonconvex MINLPs Solving nonconvex projection step NLP via a multistart heuristic using local NLP solvers Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

8 Alternating Direction Methods min x,y f (x, y) s.t. x X, y Y, g(x, y) = 0, h(x, y) 0 Choose initial values (x 0, y 0 ) X Y. for k = 0, 1,... do Compute x k+1 argmin x {f (x, y k ) : g(x, y k ) = 0, h(x, y k ) 0, x X } Compute y k+1 argmin y {f (x k+1, y) : g(x k+1, y) = 0, h(x k+1, y) 0, y Y } Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

9 ADMs: Convergence Theory A point (x, y ) Ω is called a partial minimum if holds. f (x, y ) f (x, y ) for all (x, y ) Ω, f (x, y ) f (x, y) for all (x, y) Ω Theorem (see e.g. Gorski et al. 2007) Let f, g, h be continuous, X, Y non-empty, compact, and disjoint and (x k, y k ) k=0 a sequence generated by ADM. If the solution of one optimization problem is always unique, then every convergent subsequence of { (x k, y k ) } k=0 converges to a partial minimum and the objective values of all these limit points are equal. Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

10 Idealized Feasibility Pumps are ADMs Consider the MIP min x s.t. c x Ax b x i Z for all i I Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

11 Idealized Feasibility Pumps are ADMs Consider the MIP min x s.t. c x Ax b x i Z for all i I Duplicate variables x I using the new variable vector y {0, 1} I min x,y c x s.t. x X := {x R n : Ax b, x I [0, 1] I } y Y := {0, 1} I, g(x, y) = x I y = 0 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

12 Idealized Feasibility Pumps are ADMs Consider the MIP min x s.t. c x Ax b x i Z for all i I Duplicate variables x I using the new variable vector y {0, 1} I min x,y c x s.t. x X := {x R n : Ax b, x I [0, 1] I } y Y := {0, 1} I, g(x, y) = x I y = 0 l 1 penalization of coupling condition y = x I min x,y x I y 1 s.t. x X := {x R n : Ax b, x I [0, 1] I } y Y := {0, 1} I Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

13 ADM Theory for Convex MINLP Feasibility Pumps Lemma (Geißler, Morsi, LS, Schmidt (2016)) The ADM does not cycle... and so does not the idealized feasibility pump. Theorem (Geißler, Morsi, LS, Schmidt (2016)) The idealized feasibility pump for convex MINLPs is equivalent to the ADM algorithm applied to the reformulated problem above. Thus, it terminates at a partial minimum (x, y ) after a finite number of iterations. If this partial minimum has objective value x I y 1 = 0, the point (x, y ) is feasible for the original problem. Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

14 ADM Theory for Feasibility Pumps Positive Case Idealized feasibility pump converges to a MI(NL)P-feasible partial minimum of the reformulated problem. Negative Case Idealized feasibility pump converges to a partial minimum of the reformulated problem that is not MI(NL)P-feasible. Random restarts can be seen as an attempt to escape MI(NL)P-infeasible partial minima Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

15 Another way to escape infeasibility... Penalty methods l 1 penalty function with φ 1 (x, y; µ, ρ) := f (x, y) + [α] := m µ i g i (x, y) + i=1 { 0, if α 0 α, if α < 0 µ = (µ i ) m i=1, ρ = (ρ i) p i=1 0: penalty parameters Penalty problem p ρ i [h i (x, y)] i=1 min x,y φ 1 (x, y; µ, ρ) s.t. x X, y Y Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

16 The l 1 Penalty Alternating Direction Method Choose initial values (x 0,0, y 0,0 ) X Y and penalty parameters µ 0, ρ 0 0 for k = 0, 1,... do l 0 while (x k,l, y k,l ) is not a partial minimum of the penalty problem with µ = µ k and ρ = ρ k do Compute x k,l+1 argmin x {φ 1 (x, y k,l ; µ k, ρ k ) : x X } Compute y k,l+1 argmin y {φ 1 (x k,l+1, y; µ k, ρ k ) : y Y } l l + 1 Choose new penalty parameters µ k+1 µ k and ρ k+1 ρ k Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

17 Penalty ADM: Convergence Theory Weighted l 1 feasibility measure m χ µ,ρ (x, y) := µ i g i (x, y) + i=1 p ρ i [h i (x, y)] i=1 Theorem (Geißler, Morsi, LS, Schmidt (2016)) Suppose that the assumptions hold and that µ k i for all i = 1,..., m and ρ k i for all i = 1,..., p. Moreover, let (x k, y k ) be a sequence of partial minima of the penalty problems (for µ = µ k and ρ = ρ k ) generated by PADM with (x k, y k ) (x, y ). Then there exist weights µ, ρ 0 such that (x, y ) is a partial minimizer of the feasibility measure χ µ, ρ. If (x, y ) is feasible for the original problem, then (x, y ) is a partial minimum of the original problem. The latter case can be improved if more regularity of the problem is assumed. Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

18 ADM-Exactness of l 1 Penalty Functions Theorem Let (x, y ) be a partial minimizer of min x,y f (x, y) s.t. g(x, y) = 0, x X, y Y, (1) and suppose that the Assumptions [...] hold. Then there exists a constant µ > 0 such that (x, y ) is a partial minimizer of for all µ µ and min x,y φ 1 (x, y; µ) s.t. x X, y Y φ 1 (x, y; µ) := f (x, y) + m µ i g i (x, y). i=1 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

19 Back to Feasibility Pumps Take the MINLP and duplicate the integer components x I of x The sets min x,y are compact Additional equality constraints Apply penalty ADM algorithm f (x) s.t. h(x) 0, x I = y, y Z I [l I, u I ] X := {x : h(x) 0}, Y := Z I [l I, u I ] g(x, y) = x I y = 0 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

20 Mixed-Integer Linear Problems Computational Setup C++ implementation; compiled with gcc using flag o3 LP solver: Gurobi Performance profiles (Dolan, Moré 2002) Running time; time limit 1 h Performance measure: primal-dual gap gap = Test instances: MIPLIB 2003, 2010 b p b d inf{ z : z [b d, b p ]} Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

21 Mixed-Integer Linear Problems PADM w/o local branching compared to OFP by Achterberg, Berthold inc m inc m, lb inc a inc a, lb OFP 1e 02 1e+01 1e+04 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

22 Mixed-Integer Nonlinear Problems Computational Setup C++ implementation using the GAMS Expert-Level API GAMS NLP solver: CONOPT 3.17A Test instances: MINLPLib and MINLPLib2 Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

23 Mixed-Integer Nonlinear Problems e 02 1e+01 1e+04 Red: penalty ADM based feasibility pump Blue: six feasibility pump variants of D Ambrosio et al. (2012) Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

24 Mixed-Integer Nonlinear Problems e 02 1e+01 1e+04 Red: penalty ADM based feasibility pump Blue: feasibility pump variants by Berthold (2014) Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

25 Summary Idealized feasibility pumps are alternating direction methods Convergence towards partial minima of a reformulated problem Random restarts: attempt to escape MI(NL)P-infeasible partial minima Random restarts penalty framework New penalty ADM with convergence theory Very encouraging numerical results Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,

26 Geißler, Morsi, Schewe, Schmidt (2016): Penalty Alternating Direction Methods for Mixed-Integer Optimization: A New View on Feasibility Pumps Thanks! Lars Schewe FAU Erlangen-Nürnberg Feasibility Pumps as Penalty ADMs Aussois,