1. 1 Attractive Mathematical Representations Of Decision Problems Warren Adams 11/04/03

2. 2 Research Interests Design and implementation of solution strategies for difficult (nonconvex) decision problems. Theoretical development. Algorithmic design. Computer implementation.

3. 3 Significance & Impact This talk summarizes a new, powerful procedure for constructing attractive formulations of optimization problems. The formulations generalize dozens of published papers. Striking computational successes have been realized on various problem types.

4. 4 Formulation Can Matter! Although more than one mathematical representation can accurately depict the same physical scenario, the choice of formulation can critically affect the success of solution strategies. What is an attractive formulation? How to obtain an attractive formulation?

5. 5 What Is An Attractive Formulation? Since linear programming relaxations are often used to approximate difficult problems, formulations that have tight continuous relaxations are desirable.

6. 6 Fixed Charge Network Flow (A classic example) 1 shipment cost 2 6 1 2 1 fixed cost 12 1 2 3 2 1 supply 12 demand 3 3 6 14 8

7. 7 Standard Representation

8. 8 Optimal relaxed value = 24.5. x 1 =1/4 3 3 1 6 1 shipment cost 2 6 1 2 fixed cost 12 1 2 3 2 1 supply 12 demand 3 3 6 14 8 x 2 =3/4

9. 9 Enhanced Representation

10. 10 Enhanced Representation Optimal relaxed value =29. x 1 =1 3 3 1 6 1 shipment cost 2 6 1 2 fixed cost 12 1 2 3 2 1 supply 12 demand 3 3 6 14 8 x 2 =0

11. 11 In General, How To Obtain Attractive Formulations? Attractive formulations for special problem classes can be found in the literature, but no general (encompassing) schemes exist.

12. 12 A New Perspective Historic reasoning. Convert to linear form, making any needed substitutions and/or transformations. Avoid nonlinearities. Newer reasoning. Construct nonlinearities. Then convert to linear form, using the nonlinearities to yield superior representations.

13. 13 A Method For Obtaining Attractive Formulations Reformulate the problem by incorporating additional variables and nonlinear restrictions that are redundant in the original program, but not in the relaxed version. Linearize the resulting program to obtain the problem in a different variable space.

14. 14 Reformulation-Linearization Technique (RLT) minimize c t x + d t y subject to Ax + By >= b 0=< x =<1 x binary y >= 0

15. 15 RLT: A General Approach To Attractive Formulations (Level-1) Reformulation. Multiply each constraint by product factors consisting of every 0-1 variable x i and its complement 1- x i. Apply the binary identity x i x i = x i for each i. Linearization. Substitute, for each (i,j) with i<j, a continuous variable w ij for every occurrence of x i x j or x j x i, and, for each (j,k), a continuous variable v jk for every occurrence of x j y k.

16. 16 Linearized Problem (Level-1) minimize c t x + d t y subject to Ax + By + Dw +Ev >= b x binary y >= 0 The linearized problem is equivalent to the original program in that for any feasible solution to one problem, there is a feasible solution to the other problem with the same objective value.

17. 17 Relaxation Strength? The weakest level-1 representations tend to dominate alternate formulations available in the literature, even for select problems having highly-specialized structure! As a result, we have been able to solve larger problems than previously possible.

18. 18 A Hierarchy Of Relaxations By changing the product factors, an n+1 hierarchy of relaxations emerges, with each level at least as tight as the previous level, and with an explicit algebraic characterization of the convex hull available at the highest level.

19. 19 Level-0 Representation x 1 >=0 x 2 >=0 x 2 <=1 x 1 <=1 2x 1 +2x 2 <=3 (0, 0) (1, 0) (0, 1) (1/2, 1) (1, 1/2) x2x2 x1x1

20. 20 Level-1 Representation 0.5x 1 +x 2 <=1 (2/3, 2/3) x 1 +0.5x 2 <=1 x 1 >=0 x 2 >=0 (0, 0) (1, 0) (0, 1) x2x2 x1x1

21. 21 Level-2 Representation x 1 +x 2 <=1 x 1 >=0 x 2 >=0 (0, 0) (1, 0) (0, 1) x2x2 x1x1

22. 22 Case Study: Quadratic 0-1 Knapsack Problem minimize c t x + x t Dx subject to a t x<=b x binary Capital budgeting problems. Approximates related problems.

23. 23 Computational Flavor Problem SizeClassic Formulation Level-1 Formulation NodesCPU TimeNodes CPU Time 10 0 0 8 0 20 45 0 44 0 30 421 0 102 0 40 3,899 2 826 1 50 7,043 4 771 1 60 146,430 119 2,559 3 70 92,967 99 4,465 5 80 1,232,794 1,519 8,676 9 90 **** **** 57,730 73 100 **** **** 59,001 94 Averages of ten problems solved using CPLEX 8.0. **** Average solution time exceeded the 35,000 CPU second limit.

24. 24 Computational Successes Electric Distribution System Design. Reliable Water Distribution Networks. Engineering and Chemical Process Design Problems. Time-Dynamic Power Distribution. Water Resources Management. Quadratic Assignment Problem. Capital Budgeting Problems.

25. 25 Ongoing Research Discrete variable problems. Generalizing the product factors to Lagrange interpolating polynomials. Balancing problem size and relaxation strength. Generating new families of inequalities. Applying functional product factors.

26. 26 Research Needs Wish to conduct collaborative, interdisciplinary research that blends these optimization tools with decision problems arising in electric power systems. Eager for discussions!