Optimal Decision Rules for Adjustable Robust Optimization via the lens of Fourier-Motzkin Elimination Jianzhe Zhen 1 Dick den Hertog 1 2 1 Melvyn Sim 2 Tilburg University National University of Singapore CMS, Bergamo May 31, 2017 1 / 17 Overview 1 Fourier-Motzkin elimination (FME) approach for adjustable robust optimization (ARO) problems. 2 Characterization of optimal decision rules (ODRs) for the primal and dual formulations. 3 LP-based procedure to remove redundant constraints for ARO problems. 4 Numerical experiments on the primal and dual formulations. 2 / 17 Problem Definition A two-stage ARO problem: min xPX,y s.t. c1 x Apzqx ` Bypzq ě dpzq I1 ,N2 yPR @z P W pP q , N1 1 X Ď RN1 , e.g., X “ RN ` , X “ Z A and d are affine in z z P RI1 resides in a general uncertainty set W B P RM ˆN2 is constant (a.k.a., fixed recourse) RI1 ,N2 denotes the space of functions from RI1 to RN2 that are bounded on compact sets. 3 / 17 Definition of Feasible Set Definition (Feasible set for two-stage ARO problems) The feasible set of the here-and-now decisions x in pP q is: ! ) X “ x P X | Dy P RI1 ,N2 : Apzqx ` Bypzq ě dpzq @z P W . Goal: eliminate ypzq via FME. 4 / 17 Fourier-Motzkin Elimination for Two-stage ARO problems 1 Rewrite each constraint in X in the form: Dy P RI1 ,N2 : ÿ ÿ bi1 y1 pzq ě di pzq´ aij pzqxj ´ bij yj pzq @z P W jPrN1 s @i P rM s; jPrN2 szt1u if bi1 ‰ 0, divide both sides by bi1 , we have: Dy P RI1 ,N2 : y1 pzq ě fi pzq ` g 1i pzqx ` h1i y zt1u pzq y1 pzq ď 0ě fj pzq ` g 1j pzqx ` h1j y zt1u pzq fk pzq ` g 1k pzqx ` h1k y zt1u pzq @z P W if bi1 ą 0, @z P W if bj1 ă 0, @z P W if bk1 “ 0. 5 / 17 Fourier-Motzkin Elimination for Two-stage ARO problems 1 Rewrite each constraint in X in the form: Dy P RI1 ,N2 : ÿ ÿ bi1 y1 pzq ě di pzq´ aij pzqxj ´ bij yj pzq @z P W jPrN1 s @i P rM s; jPrN2 szt1u if bi1 ‰ 0, divide both sides by bi1 , we have: Dy P RI1 ,N2 : y1 pzq ě fi pzq ` g 1i pzqx ` h1i y zt1u pzq y1 pzq ď 0ě 2 fj pzq ` g 1j pzqx ` h1j y zt1u pzq fk pzq ` g 1k pzqx ` h1k y zt1u pzq @z P W if bi1 ą 0, @z P W if bj1 ă 0, @z P W if bk1 “ 0. Let Xzt1u be the feasible set after y1 is eliminated, and defined by: Dy zt1u P RI1 ,N2 ´1 : fj pzq ` g 1j pzqx ` h1j y zt1u pzq ě fi pzq ` g 1i pzqx ` h1i y zt1u pzq 0 ě fk pzq ` Complexity: OpM 2 N2 g 1k pzqx ` h1k y zt1u pzq @z P W if bj1 ă 0 and bi1 ą 0, @z P W if bk1 “ 0. q. 5 / 17 Simple Example — 1 Example (Eliminate y1 via FME) We eliminate y1 in the following linear inequalities via FME: y1 pzq ´ z 1 x ď 0 @z P W y1 pzq ´ y2 pzq ď 0 @z P W ´y1 pzq ` x3 ` 2y2 pzq ď 0 @z P W ´y2 pzq ď z2 Step 1. @z P W. Step 2. 1 y1 pzq ď z x @z P W x3 ` 2y2 pzq ď z 1 x @z P W y1 pzq ď y2 pzq @z P W x3 ` 2y2 pzq ď y2 pzq @z P W x3 ` 2y2 pzq ď y1 pzq @z P W ´y2 pzq ď z2 @z P W. ´y2 pzq ď z2 @z P W. 6 / 17 Simple Example — 2 Example (Eliminate y2 via FME) We now also eliminate y2 : x3 ` 2y2 pzq ď z 1 x x3 ` y2 pzq ď 0 ´y2 pzq ď z2 Step 1. 1 y2 pzq ď pz 1 x ´ x3 q 2 y2 pzq ď ´x3 ´z2 ď y2 pzq @z P W @z P W @z P W. Step 2. @z P W @z P W ´2z2 ď z 1 x ´ x3 ´z2 ď ´x3 @z P W @z P W. @z P W. Linear decision rule is optimal for y2 pzq. 7 / 17 Optimal Decision Rules — General Theorem There exist ODRs that are piecewise affine functions for y in pP q. This theorem holds for Problem pP q with uncertain parameters that: appear on both sides of the constraints reside in a general uncertainty set 8 / 17 Optimal Decision Rules — General Theorem There exist ODRs that are piecewise affine functions for y in pP q. This theorem holds for Problem pP q with uncertain parameters that: appear on both sides of the constraints reside in a general uncertainty set which generalizes the existing result of [Bemporad et al.(2003)] stimulates a generalization of the recent methods in: [Bertsimas and Georghiou (2015)] [Ben-Tal et al.(2016)]. 8 / 17 Dual Feasible Set [Bertsimas and de Ruiter (2016)] Suppose pP q has a polyhedral uncertainty set ! ) Wpoly “ z P RI1 | Dv P RI2 : P 1 z ď ρ . Remember: ! ) X “ x P X | Dy P RI1 ,N2 : Apzqx ` Bypzq ě dpzq @z P Wpoly . 9 / 17 Dual Feasible Set [Bertsimas and de Ruiter (2016)] Suppose pP q has a polyhedral uncertainty set ! ) Wpoly “ z P RI1 | Dv P RI2 : P 1 z ď ρ . Remember: ! ) X “ x P X | Dy P RI1 ,N2 : Apzqx ` Bypzq ě dpzq @z P Wpoly . The dual feasible set: $ ω 1 pA0 x ´ d0 q ě ρ1 λpωq & D X “ x P X Dλ P RM,K : p1i λpωq “ pdi ´ Ai xq1 ω % λpωq ě 0 where , @ω P U . @ω P U, @i P rI1 s @ω P U ! ) 1 Upoly “ ω P RM ` | B ω “ 0 . Theorem ([Bertsimas and de Ruiter (2016)]) X “ X D. 9 / 17 Optimal Decision Rules — Simplex Suppose ! Wsimplex “ z P RI`1 ) 11 z ď 1 . Theorem There exist LDRs that are optimal for y in (P) with Wsimplex . 10 / 17 Optimal Decision Rules — Simplex Suppose ! Wsimplex “ z P RI`1 ) 11 z ď 1 . Theorem There exist LDRs that are optimal for y in (P) with Wsimplex . “Proof”: " D X “ xPX M,1 Dλ P R ω 1 pA0 x ´ d0 q ě λpωq ` ˘` : λpωq ě pdi ´ Ai xq1 ω @ω P U @ω P U, @i P rI1 s * . This coincides with the recent finding in [Ben-Ameur et al.(2016), Corollary 2], and generalizes [Bertsimas and Goyal (2012), Theorem 1]. Theorem ([Z. and den Hertog (2015)]) There exist polynomials of (at most) degree N1 and linear in each zi , i P rI1 s, that are optimal decision rules for y in pP q. 10 / 17 Optimal Decision Rules — Box Suppose ! Wbox “ z P RI1 ) lďzďu . Theorem ` ˘` are ODRs for the The convex two-piecewise affine functions pdi ´ Ai xq1 ω adjustable variables λi , i P rI1 s, in (D). “Proof”: " XD “ x P X Dλ P RM,I1 : ¨ ¨ ¨ ě pu ´ lq1 λpωq ` ˘` λi pωq ě pdi ´ Ai xq1 ω @ω P U @ω P U, @i P rI1 s * . 11 / 17 Redundant Constraint Identification Suppose pP q has only rhs uncertainty, i.e., ! ) X “ x P X | Dy P RI1 ,N2 : Ax ` Bypzq ě dpzq @z P W . Theorem The l-th constraint is redundant in X if the following optimization problem Zl: “ min a1l x ` b1l y ´ dl pzq s.t. a1i x ` b1i y ě di pzq x,y,z N1 z P W, x P R ,y P R has a nonnegative optimal objective value, i.e., pě 0 ñ Redundant!q @i P rM sztlu N2 . Zl: ě 0. 12 / 17 Inner Approximation via Decision Rules Impose decision rules F I1 ,N2 ´|S| Ď RI1 ,N2 ´|S| on y zS : ! ) XpzS “ x P X|Dy zS P F I1 ,N2 ´|S| : Gpzqx ` Hy zS pzq ě f pzq @z P W ! ) “ x P X|Dy zS P F I1 ,N2 ´|S| , y S P RI1 ,|S| : Apzqx ` Bypzq ě dpzq @z P W . Theorem Xp Ď XpzS Ď XzrN2 s “ X , for S Ď rN2 s. 13 / 17 Numerical Example — A Two-stage RO Problem Example (Lot-sizing on a network) min xPX,yij ,τ s.t. c1 x ` τ ÿ tij yij pzq ď τ @z P W i,jPrN s xi ` ÿ yji pzq ´ jPrN s ÿ yij pzq ě zi @z P W, i P rN s jPrN s yij pzq ě 0, yij P RN,1 @z P W, i, j P rN s. ci : storage cost of store i x P X: stock allocations in X z P W: the uncertain demands reside in a budget uncertainty set W tij : transportation cost per unit from store i to store j yij : units of good transported from store i to store j. 14 / 17 Numerical Results — Primal v.s. Dual Dimensions of y and λ are N 2 and N ` 1, respectively. 15 / 17 Numerical Results — Primal v.s. Dual Dimensions of y and λ are N 2 and N ` 1, respectively. Table: Lot-sizing on a Network for N P t5, 10u. N=5 P D N=10 P D #Elim. RCI Gap% TTime #Elim. FME Gap% Time #Elim. RCI Gap% TTime #Elim. FME Gap% Time 1 30 3.3 0.1 1 11 3.3 0.1 1 110 6.0 0.1 1 21 6 0.1 11 37 2.9 12.9 2 10 2.8 0.1 12 100 6.0 14.8 5 43 4.6 0.2 15 75 1.7 58.3 3 13 2.3 0.1 17 133 6.0 52.6 7 135 2.6 0.4 19 101 0.7 223.2 4 19 1 0.1 19 180 5.9 100.7 8 261 1.8 0.9 22 116 0.1 394.3 5 33 0.2 0.1 21 276 5.8 987.7 9 515 0.8 1.9 25 127 0 550.3 6 272 0 0.1 22 343 5.7 1639.8 10 1025 0.2 4.6 – – – – – – 100 * * * 11 149424 * * 15 / 17 Numerical Results — Dual Table: Lot-sizing on a Network for N P t15, 20, 30u. N=15 D N=20 D N=30 D #Elim. FME Red.% Time FME Red.% Time FME Red.% Time Red.%“ 1 31 0 0.3 41 0 0.6 61 0 2.2 2 31 -0.1 0.3 41 -0.1 0.5 61 0 2.2 sol.´LDR LDR 3 33 -0.4 0.3 43 -0.2 0.6 63 -0.1 2.6 4 39 -0.5 0.4 49 -0.3 0.8 69 -0.2 3.4 5 53 -0.7 0.5 63 -0.4 1.3 83 -0.2 5.8 6 83 -0.9 1 93 -0.5 2.5 113 -0.3 15.0 7 145 -1.6 1.5 155 -0.9 3.5 175 -0.4 54.6 8 271 -1.9 3.6 281 -1 10.1 301 -0.5 55.7 9 525 -2.2 8.9 535 -1.2 36.1 555 -0.6 214.5 10 1035 -2.8 25.2 1045 -1.5 67.7 1065 -0.8 522.2 11 2057 -3.4 125.1 2067 -1.8 206.2 2087 * * ˆ 100%. 16 / 17 Conclusions & Future Research FME enables us to gain new insights on ODRs successively improve the solutions obtained from the existing methods. FME cannot be applied to non-fixed recourse Bpzq integer recourse decisions ypzq P ZN2 . Future research: characterize ODRs for multistage problems adapt the existing procedures to improve the computability of FME approach. Zhen, J., D. den Hertog, M. Sim. (2016) Adjustable Robust Optimization via Fourier-Motzkin Elimination. Available on Optimization Online. 17 / 17 Bemporad, A., F. Borrelli, M. Morari (2003) Min-max control of constrained uncertain discrete-time linear systems. IEEE Transactions Automatic Control 48(9):1600–1606. Ben-Ameur, W., G. Wang, A. Ouorou, M. Zotkiewicz (2016) Multipolar Robust Optimization. Available at: arxiv.org/pdf/1604.01813.pdf. Ben-Tal, A., O. El Housni, V. Goyal (2016) A tractable approach for designing piecewise affine policies in dynamic robust optimization. Available at: optimization-online.org/DB_FILE/2016/07/5557.pdf. Bertsimas, D., F. de Ruiter (2016) Duality in two-stage adaptive linear optimization: faster computation and stronger bounds. INFORMS Journal on Computing 28(3):500–511. Bertsimas, D., V. Goyal (2012) On the power and limitations of affine policies in two-stage adaptive optimization. Mathematical Programming Ser. A 134(2):491–531. Bertsimas, D., A. Georghiou (2015) Design of near optimal decision rules in multistage adaptive mixed-integer optimization. Operations Research 63(3):610–627. 17 / 17 Bertsimas, D., M. Sim, M. Zhang (2016) Distributionally adaptive optimization. Available at: www.optimization-online.org/DB_FILE/2016/03/5353.pdf. Bertsimas, D. J. Tsitsiklis (1997) Introduction to Linear Optimization. Athena Scientific. Caron, R., J. McDonald, C. Ponic (1989) A degenerate extreme point strategy for the classification of linear constraints as redundant or necessary. Journal of Optimization Theory and Applications 62:225–237. J. Fourier (1826) Reported in: Analyse des travaux de l’Académie Royale des Sciences, pendant l’année 1824, Partie mathématique. Histoire de l’Academie Royale des Sciences de l’Institut de France 7:47–55. T. Motzkin (1936) Beiträge zur Theorie der linearen Ungleichungen, University Basel Dissertation. Jerusalem, Israel. Zhen, J. and D. den Hertog (2015) Computing the maximum volume inscribed ellipsoid of a polytopic projection. INFORMS Journal on Computing, to appear (available at: optimization-online.org/DB_FILE/2015/01/4749.pdf). 17 / 17
© Copyright 2026 Paperzz