# Strong Bounds for Linear Programs with Cardinality Limited

- Slides: 28

Strong Bounds for Linear Programs with Cardinality Limited Violation (CLV) Constraint Systems Ali Tuncel Ronald L. Rardin Purdue University [email protected] edu University of Arkansas [email protected] edu Jean-Philippe Richard Purdue University [email protected] edu Mark Langer, M. D. Indiana University [email protected] edu Workshop on Mixed Integer Programming August 4 -7, 2008

LP’s with CLV Constraints l Models LPCLV have all constraints and the objective linear, but some constraints belong to systems where up to k of m may be violated

MIP Formulation of CLV Constraints l If some outer limit on violation CLV system is easily modeled is available, the l But a difficult MILP often results, especially when there is a great difference between and

Application: Portfolio Optimization l LPCLV models arise in Value at Risk portfolio optimization, where decision variables xj represent the investment from budget P in asset j , and CLV constraints reflect risk under various scenarios i, at most k of which may exceed b

Application: Radiation Therapy (Fluence) Planning • Radiation is delivered by an accelerator that rotates 360 degrees around the patient • Full beam area is large, around 10 cm square. However, Intensity Modulated Radiation Therapy (IMRT) can deliver shaped intensity maps composed of 3 -10 mm rectangular beamlets with independent fluence

Application: Radiation Therapy (Fluence) Planning • The goal of intensity/fluence/timeon planning optimization is to assign optimal beamlet intensities xj so that prescribed doses are delivered to different types of tissues: l Primary Target: Tumor region. Maximize dose within homogeneity requirements l Secondary Targets: Minimum dose limits to eliminate risk of microscopic infection. l Healthy/Normal Tissues: Maximum dose limits on all or specified fractions of surrounding tissues to avoid complications. l Dose-volume limits are ones on fractions of normal tissue volume (e. g. no more than 20% of points in the tissue can receive a dose exceeding 60)

Dose-Volume Tissues l Tissues are modeled as collections of imbedded points i + + + l Dose at any point is approximately linear in beamlet fluence xj l Leads to a CLV system for each dose-volume constrained tissue with other constraints linear + + +

Computational Challenge – Hard Cases The difference between b and b values influences the initial optimality gap.

Rounding Heuristic l An LP-rounding heuristic often gives feasible integer solutions 1. 2. 3. 4. l Solve the LP relaxation Sort the points in each dose-volume tissue by LP relaxation dose Force all but the highest-dose k to satisfy b Solve the modified LP relaxation Results are good if the LP is a fair to good approximation of the MIP

Computational Challenge Upper Bound 116. 61 Lower Bound 80. 62 Gap 45% CPU Secs 86 Lung 3 A 110. 01 Lung 3 B 104. 19 CLung. A 90. 07 CLung. B 92. 89 42. 98* 41. 03* 58. 40 59. 95 156%* 154%* 54% 55% 689 1012 530 184 Case Lung 1 LP relaxation and rounding heuristic * feasible solution from later methods Compared to CPLEX 11. 0 Case Lung 1 Lung 3 A Lung 3 B CLung. A CLung. B Upper Lower Bound 115. 93 81. 38 109. 99 37. 83 104. 18 37. 99 96. 62 58. 24 90. 42 59. 93 CPU B & B Gap Secs Nodes 42% 86400 48786 191% 86400 155 174% 86400 140 66% 86400 120 51% 86400 1220

Disjunctive Cutting Planes

LPCLVs As Disjunctive Programs l m=5, K=2 x 2 l b = 20 x 1 Disjunctive programming studies optimization problems over disjunctions (unions) of polyhedra (LP feasible sets) Easy to see how LPCLVs can be viewed this way as a union over all the LPs with different sets of (m-k) of the tighter b RHSs explicitly enforced (here all 3 of 5)

Valid Inequalities for Disjunctive Programming l Balas proved that all needed valid inequalities for disjunctions of polyhedra can be constructed by rolling up each of the member linear systems with nonnegative multipliers, and picking the lowest of the rolled coefficients on the LHS and the largest on the RHS

Valid Inequalities for LPCLV – Main Disjunctive Inequality x 2 m=5, K=2 b = 20 Average of lowest 3 values Main Disjunctive Inequality x 1

Valid Inequalities for LPCLV – A Family of Disjunctive Inequalities Lemma. For any subset S of i=1, …, m with K or more elements, the implied LPCLV(A[S], b, K) is a relaxation of the full LPCLV(A, b, K) where A[S] is the row sub- matrix of A for rows in S Proposition. For any subset S of i=1, …, m with more than K elements, the inequality is valid for full system LPCLV(A, b, K), where each is the average of the (|S|-K) smallest coefficients in rows in S.

Valid Inequalities for LPCLV – Disjunctive Support Inequalities Proposition. Let subset Sk of i=1, …, m be the K+1 rows with largest coefficients on xk. Then the inequality is valid for full system LPCLV(A, b, K), where each is smallest j coefficient for rows in Sk. Furthermore the inequality supports the convex hull of solutions to the full x 2 LPCLV(A, b, K) (at xk = b / , other xj = 0) m=5, K=2 b = 20 Average of the lowest value Disjunctive Support Inequality x 1

Results for Disjunctive Cuts Upper Bound 116. 61 Lower Bound 80. 62 Gap 45% CPU Secs 86 Lung 3 A 110. 01 Lung 3 B 104. 19 CLung. A 90. 07 CLung. B 92. 89 42. 98* 41. 03* 58. 40 59. 95 156%* 154%* 54% 55% 689 1012 530 184 Case Lung 1 * feasible solution from later methods Case Lung 1 LP relaxation with disjunctive cuts and rounding heuristic LP relaxation and rounding heuristic (except *) Upper Lower Bound Gap 94. 24 81. 63 15% CPU Secs # of Cuts 56 1060 Lung 3 A 62. 12 42. 98 45% 600 2508 Lung 3 B 62. 26 41. 03 52% 645 2512 CLung. A 90. 07 58. 40 54% 540 890 CLung. B 92. 89 59. 95 55% 205 898

Single Constraint Enumeration

SCE Concept l Consider relaxations enforcing only one row of the CLV system at a time

SCE Concept Proposition. Let zi denote the optimal solution value of SCEi. Then the k+1 st smallest such zi is a valid upper bound on the optimal solution value of the full LPCLV. (Proof idea: There can be at most k of the single constraints violated in LPCLV, but thereafter all constraints must be satisfied at the b value. ) l Row i zi 3 43 8 65 1 78 6 52 4 67 5 81 9 56 2 72 7 90 Example with m=9, k=3 Direct SCE (i) solves all one-row problems, (ii) sorts solution values, and (iii) picks the k+1 st smallest

Adaptive Single Constraint Enumeration (ASCE) l Do better by using one-row problems solved to get adaptive upper/lower bounds for unsolved SCEi K+1 st Lowest Upper Bound Is SCE Upper Bound … … … Solved Case Each Box Shows SCE Value Upper – Lower Range with Bar for True Optimal Value Bounds Improved by Solving Other SCEs K+1 st Lower Bound Is SCE Lower Bound

ASCE – Global Bounds l The K+1 st lowest of the upper bounds is a running upper bound on the ultimate SCE bound obtained (max of these upper bounds >= max of their z’s >= max of ultimate K+1 z’s) l The K+1 st lowest of the lower bounds is a running lower bound on the ultimate SCE bound obtained (max of these lower bounds <= max of lower bounds for ultimate K+1 <= max of their z’s)

ASCE – Surrogate Relaxation Upper Bounds l Each SCEi solved produces dual multipliers that can be used to quickly update running upper bounds on solution values for other SCEq l l l Construct a surrogate constraint valid for each SCEq using the dual solution for SCEi without any multipliers on the CLV constraints to roll up all the others Gx<=h For each q not i in turn, solve a simple relaxation of SCEq having only such surrogate constraints and the single CLV inequality for q These yield upper bounds on the solution values of the corresponding SCEq , and the running best upper bounds on their values can be updated

ASCE – Common Feasible for Lower Bounds l If an SCEi solution vector is also feasible for constraint q of the CLV system, then i’s solution value is a lower bound on that SCEq and the running lower bound on that solution value can be updated

ASCE - Pruning l l If the running upper bound on the value of SCEq is already < the best known LPCLV feasible solution value, then q is certain to be among the lowest K+1 that establish the global bound. No further processing of SCEq is needed. If at least K+1 of the running upper bounds are no greater that the running lower bound for SCEq then it cannot be in the lowest K+1. No further exploration of q is required

Results with SCE and ASCE Case Lung 1 Upper Lower Bound Gap 94. 24 81. 63 15% CPU Secs # of Cuts 56 1060 Lung 3 A 62. 12 42. 98 45% 600 2508 Lung 3 B 62. 26 41. 03 52% 645 2512 CLung. A 90. 07 58. 40 54% 540 890 CLung. B 92. 89 59. 95 55% 205 898 LP relaxation with disjunctive cuts and SCE/ASCE and rounding heuristic LP relaxation with disjunctive cuts and rounding heuristic Upper Lower Case Bound Gap Lung 1 86. 46 81. 63 6% ASCE Secs 1039 SCE Secs 3498 Lung 3 A 55. 39 42. 98 29% 14169 87006 Lung 3 B 55. 60 41. 03 36% 12632 84265 CLung. A 63. 35 58. 40 8% 5709 7628 CLung. B 67. 97 62. 91 8% 1574 1864

Conclusions l l l Linear Programming Relaxations of MIP formulations of LPCLV problems can be arbitrarily weak Commercial MIP solver (CPLEX 11) is not effective in closing the initial large optimality gaps Efficient valid inequalities based on disjunctive programming theory can be generated to: l l l Strengthen LP relaxations and significantly reduce initial optimality gaps when problems are dense Help find better feasible solutions within the rounding heuristic framework Improve run times

Conclusions l SCE procedure can be implemented as a standalone procedure or in branch and bound l l Significantly reduces upper bound values. In some cases, finds better feasible solutions l Adaptive SCE accomplishes the SCE computation in significantly reduced time l The combined algorithm utilizing disjunctive inequalities and the ASCE procedure within the rounding heuristic framework produces lower optimality gaps and better feasible solution values

- Strong Bounds for Linear Programs with Cardinality Limited
- Strong Boards Strong Plan Strong OrganizationNPI 2015 Don
- Strong AverageCase Circuit Lower Bounds from Nontrivial Derandomization
- Strong AverageCase Circuit Lower Bounds from Nontrivial Derandomization
- Lower Bounds Sorting in Linear Time Comp 122