IV.

Violations of Linear Programming Assumptions

Some types of Mathematical Programming problems violate at least one condition of strict Linearity
Deterministic Nature Additivity Direct Proportionality Fractionality

Or do not satisfy the additional conditions of:

- Nonnegativity - Single goal

We now look at how to handle such situations.

Violations of:
Deterministic Nature

Are Handled With:

Sensitivity Analysis, Parametric Programming (PP), Stochastic Programming (SP) Statistical Programming(STP) Revise/Correct Formulation Nonlinear Programming (NLP)

Additivity Direct Proportionality Fractionality Nonnegativity Single Goal

Integer Programming (IP) Creative Decision Variables Goal Programming (GP)

A. Integer Programming
- Integer Programming (IP): Branch of linear programming in which at least one decision variable is restricted to integer values. - Mixed Integer Program (MIP): - A linear program in which the values of some decision variables are limited to integer values and others are allowed to assume fractional values. - 0/1 Integer Program (Binary Program or BIP) - A linear program in which the values of decision variables are limited to binary values (zero or one). Often used for problems that involve Yes/No types of decisions. The Television Production Problem is an excellent example of a scenario for which the decision variables should be restricted to integer values - we were fortunate that this happened naturally in that problem!

Consider the following example. Riggs Paint & Supply has available 3 different processes for producing standard white house paint. Each process has a capacity and a per gallon processing cost given below: Process 1 2 3 Processing Cost/gal. $5 $4 $3 Maximum Daily Capacity (Gals.) 2000 3000 4000

Daily demand is 3500 gals. What processes should be used? How many gallons should be made by each process?

If we allow non integer quantities of paint to be produced by each process (certainly reasonable since all processes are making the same paint), then an appropriate formulation would be: minimize C= 5x1 + 4x2 + 3x3 subject to: x1 + x2 + x3 = 3500 (daily demand) 2000 (process 1 capacity) x1 x2 3000 (process 2 capacity) x3 4000 (process 3 capacity) x3 0 (nonnegativity) x1 , x2,

Where x1 is the number of gallons produced using process 1 x2 is the number of gallons produced using process 2 x3 is the number of gallons produced using process 3 What if, on the other hand, each process produced a different quality of paint?

If each process produced a different quality of paint, then we might wish to restrict production by each process to whole cans (gallons). An appropriate formulation would be: minimize C= 5x1 + 4x2 + 3x3 subject to: x1 + x2 + x3 = 3500 (daily demand) x1 2000 (process 1 capacity) x2 3000 (process 2 capacity) x3 4000 (process 3 capacity) x1 , x2, x3 0 (nonnegativity) x1 , x2, x3 are integer

Where x1 is the number of gallons produced using process 1 x2 is the number of gallons produced using process 2 x3 is the number of gallons produced using process 3 This is an Integer Programming formulation.

Now consider a fixed cost for setting up each of the processes. If the setup costs are $100 for process 1, $200 for process 2, and $300 for process 3, an appropriate formulation would be: minimize C= 5x1 + 4x2 + 3x3 + 100y1 + 200y2 + 300y3 subject to: x1 + x2 + x3 = 3500 (daily demand) 2000y1 (process 1 capacity) x1 x2 3000y2 (process 2 capacity) x3 4000y3 (process 3 capacity) x3 0 (nonnegativity) x1 , x2 , x1 , x2, x3 are integer y1 , y2, y3 are binary

Where xi is the number of gallons produced using process i yi is 1 if process i is used and 0 otherwise This is a Binary Integer Programming formulation.

- Solving Integer Programming problems graphically LP Relaxation: the version of an integer programming problem that results when the integer restrictions are ignored (relaxed). IP problems with two decision variables can be solved graphically. Consider the following LP problem: maximize = x1 + 5x2 subject to: x1 + 10x2 20 (constraint 1) x1 2 (constraint 2) x2 0 (nonnegativity) x1 , x1 , x2 are integer

Graphically, the Linear Programming relaxation problem looks like this:

x2
10

The optimal solution is at Extreme Value C (x1 = 2.0, x2= 1.8, =11.0). What would happen if we enforced the integer restrictions on x1 and x2?
x1 2

C
=1 0
x1 + 10x
2

20

10

15

x1

Why dont we simply round the LP optimal values of the decision variables (x1 =2.0, x2= 1.8)? Rounding up yields x1 = 2, x2= 2, =12 - this is a better value of the objective function, but is unfortunately infeasible. Rounding down yields x1 = 2, x2= 1, =7 - this is feasible but is unfortunately infeasible a much worse value of the objective function (can we find a better feasible solution?). What changes do the integer restrictions graphically imply for our Linear Programming relaxation problem?

x2
10

The feasible region now must lie on the lines that represent the integer restrictions on the decision variables (x1 and x2) as well as within the feasible region of the LP relaxation problem.
x1 2

X1 is integer X2 is integer
B C
x1 + 10x
2

20

10

15

x1

The feasible region is now a series of points (that were formed by the feasible lattice). x2
10

10

15

x1

x2
10

We now move the objective function in the direction of improvement until we hit the last point in the feasible region. The optimal solution is x1 = 0, x2= 2, =10 - this is both feasible and superior to the solution that resulted from rounding down the solution to the LP relaxation!

B
=1 0

10

15

x1

x2
10

If only x1 was restricted to integer values, the the feasible region for this IP problem would look like this:

x1 2

x1 + 10x

20

10

15

x1

x2
10

If only x2 was restricted to integer values, the the feasible region for this IP problem would look like this:

x1 2

x1 + 10x

20

10

15

x1

- Solving Integer Programming problems algorithmically Cutting Plane Algorithm - Set of methods for solving IP problems for which successive constraints (cuts) are added to the LP relaxation to force the extreme points of the feasible region to be at integer points. Branch & Bound Method - A cutting plane approach to solving IP problems through systematic partial enumeration. Basically: A Parent Problem (the LP relaxation with some integrality constraints enforced) is solved Then additional constraints (which cut away parts of the feasible region that violate integrality constraints) are added to form varying Descendent Linear Programs which are then solved. These become Parent Problems, and the cycle is repeated until the optimal solution to the IP is found.

The steps of the Branch & Bound Method are: 1. Solve the Parent Problem (the LP relaxation of the original Integer Programming problem in the first iteration of this algorithm). This solution provides: the current upper bound for its direct Descendents in maximization problems. the current lower bound for its direct Descendents in minimization problems. 2. Check the status of all decision variables that have integrality constraints in the original IP problem: If all integrality constraints are satisfied, the solution is feasible. If this is the best current integer solution, retain as the lower bound (for maximizations) or as the upper bound (for minimizations). Do not proceed down this branch. Proceed if any integrality constraint is violated.

3. Arbitrarily select a decision variable that has an integrality constraint in the original IP problem that is unmet by the current solution. Create two Descendent Problems by branching: Round up the current value of the chosen variable to the nearest integer and add a constraint that forces the chosen variable to be at least as great () as the result. Round down the current value of the chosen variable to the nearest integer and add a constraint that forces the chosen variable to be no more than () as the result.

4. For each descendent problem If the resulting solution does not satisfy all integrality constraints from the original IP but is superior to the lower bound (for maximizations) or the upper bound (for minimizations), continue down this path (i.e., use this as a future Parent Solution). If the solution satisfies all integrality constraints from the original IP, do not continue down this path. If this is the best current integer solution, retain this solution as the lower bound (for maximization problems) or as the upper bound (for minimization problems). If the resulting solution is not superior to the lower bound (for maximizations) or the upper bound (for minimizations), do not continue down this path.

5. Continue until all branches result in either i) a solution that is inferior to the lower bound (for maximizations) or the upper bound (for minimizations) or ii) a solution that satisfies all integrality constraints from the original IP. At this point the remaining lower bound (for maximizations) or upper bound (for minimizations) is the optimal solution.

Example - Recall the IP problem we solved graphically: maximize = x1 + 5x2 subject to: x1 + 10x2 20 (constraint 1) x1 2 (constraint 2) x1 , x2 0 (nonnegativity) x1 , x2 are integer

To solve this problem using the Branch & Bound Method, we must first solve the LP Relaxation: maximize = x1 + 5x2 subject to: x1 + 10x2 20 (constraint 1) x1 2 (constraint 2) x1 , x2 0 (nonnegativity)

The optimal solution is at Extreme Value C (x1 = 1.0, x2= 1.8, =11.0).

x2
10

Now we must systematically enforced the integer restrictions on x1 and x2. Since the value of x1 in our current solution satisfies its integrality constraint, we will branch on x2.
x1 2

C
=1 0
x1 + 10x
2

20

10

15

x1

Problem 1 x1 = 1.0 x2 = 1.8 = 11.0 x2 1.0 Problem 2 x1 = 2.0 x2 = 1.0 = 7.0 x2 2.0 Problem 3 x1 = 0.0 x2 = 2.0 = 10.0

The formulation for descendent problem 2 is: maximize = x1 + 5x2 subject to: x1 + 10x2 20 (constraint 1) x1 2 (constraint 2) x2 1 (cutting plane constraint) x2 0 (nonnegativity) x1 ,

The optimal solution is at x1 = 2.0, x2= 1.0, = 7.0. This satisfies all integrality constraints (and so is feasible), so we will no longer branch in this direction. This is also the best current solution that satisfies all integrality constraints, so it constitutes the current lower bound for our maximization problem.

The Cutting Plane for Problem 2 looks like this: x2

10 x1 2

B
=1 0
x1 + 10x
2

20

x2 1

10

15

x1

The formulation for descendent problem 3 is: maximize = x1 + 5x2 subject to: x1 + 10x2 20 (constraint 1) x1 2 (constraint 2) x2 2 (cutting plane constraint) x2 0 (nonnegativity) x1 ,

The optimal solution is at x1 = 0.0, x2= 2.0, = 10.0. This satisfies all integrality constraints (and so is feasible), so we will no longer branch in this direction. This is also now the best current solution that satisfies all integrality constraints, so it supercedes our previous lower bound and constitutes the current lower bound for our maximization problem.

The Cutting Plane for Problem 3 looks like this: x2

10

Since all branches have met our stopping criteria, we have an optimal solution to the original IP problem at x1 = 0.0, x2= 2.0, = 10.0.
x1 2 x1 2

B
=1 0
x1 + 10x
2

x2 2
20

10

15

x1

Another example - consider this IP problem: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x1 , x2 , x3, x4 x2 , x3, x4 x1 , 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (nonnegativity) (integrality)

To solve this problem using the Branch & Bound Method, we must first solve the LP Relaxation: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x1 , x2 , x3, x4 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (nonnegativity)

The optimal solution is x1= 0.00 , x2= 4.00 , x3= 1.00 , x4= 0.25, = 27.75. Now we must systematically enforced the integer restrictions on x1, x2, x3, and x4. Since the values of x1, x2, and x3 in our current solution satisfy their integrality constraints, we will branch on x4.

Problem 1 x1 = 0.00 x2 = 4.00 x3 = 1.00 x4 = 0.25 = 27.75

x4 0.0
Problem 2 x1 = 0.14 x2 = 4.00 x3 = 1.00 x4 = 0.00 = 27.14

x4 1.0
Problem 3 x1 = 0.00 x2 = 3.00 x3 = 1.00 x4 = 1.00 = 25.0

The formulation for descendent problem 2 is: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x4 x1 , x2 , x3, x4 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (cutting plane) 0 (nonnegativity)

The optimal solution is x1 = 0.14 , x2= 4.00 , x3= 1.00 , x4= 0.00, =27.14. This does not satisfy all integrality constraints, so we will continue to branch in this direction.

The formulation for descendent problem 3 is: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x4 x2 , x3, x4 x1 , 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 1 (cutting plane) 1 (nonnegativity)

The optimal solution is x1 = 0 , x2= 3 , x3= 1 , x4= 1, =25. This does satisfy all integrality constraints, so we will not branch in this direction any further. This is also the best current solution that satisfies all integrality constraints, so it constitutes the current lower bound for our maximization problem.

We continue by branching on descendent problem 2: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x4 x1 , x2 , x3, x4 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (cutting plane) 0 (nonnegativity)

The optimal solution here is x1 = 0.14 , x2= 4.00 , x3= 1.00 , x4= 0.00, =27.14, so we now branch on x1.

Problem 1 x1 = 0.00 x2 = 1.00 x3 = 4.00 x4 = 0.25 = 27.75

x4 0.0
Problem 2 x1 = 0.14 x2 = 4.00 x3 = 1.00 x4 = 0.00 = 27.14 Problem 4 x1 = 0.00 x2 = 4.00 x3 = 1.00 x4 = 0.00 = 27.00

x4 1.0
Problem 3 x1 = 0.00 x2 = 3.00 x3 = 1.00 x4 = 1.00 = 25.0 Problem 5 x1 = 1.00 x2 = 2.00 x3 = 1.00 x4 = 0.00 = 18.00

x1 0.0

x1 1.0

The formulation for descendent problem 4 is: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x4 x1 x2 , x3, x4 x1 , 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (cutting plane) 0 (cutting plane) 0 (nonnegativity)

The optimal solution is x1 = 0 , x2= 4 , x3= 1 , x4= 0, =27. This does satisfy all integrality constraints, so we will not branch in this direction any further. This is also the best current solution that satisfies all integrality constraints, so it now constitutes the current lower bound for our maximization problem.

The formulation for descendent problem 5 is: maximize = x1 + 5x2 + 7x3 + 3x4 subject to: 7x1 + 3x2 + 2x3 + 4x4 8x1 + 2x2 + 3x3 + 5x4 x1 x2 x3 x4 x4 x1 x1 , x2 , x3, x4 15 (constraint 1) 17 (constraint 2) 4 (constraint 3) 4 (constraint 4) 1 (constraint 5) 1 (constraint 6) 0 (cutting plane) 1 (cutting plane) 0 (nonnegativity)

The optimal solution is x1 = 1 , x2= 2 , x3= 1 , x4= 0, =18. This does satisfy all integrality constraints, so we will not branch in this direction any further. However, this solution is inferior to our current lower bound (=27 at problem 5) and so this solution is terminal (and we abandon our search down this path).

Since all branches have ended in solutions that satisfy all integrality constraints (or have feasible solutions worse than the current lower bound) we are finished branching. The optimal solution to the original IP problem is the solution for descendent problem 5: x1 = 0 x2 = 4 x3 = 1 x4 = 0 = 27

- Uses for Integer Restricted Variables: Either-Or Variables - Suppose there is a choice as to which of two projects that can occur, so that it is necessary that at most one of two corresponding DVs (say x1 or x2) is equal to one. This could be accomplished by incorporating the following constraint into the model: x1 + x2 1 x1, x2 = 0 or 1 If it is necessary that exactly one of two projects must be chosen, this could be accomplished by incorporating the following constraint into the model: x1 + x2 = 1 x1, x2 = 0 or 1

k out of r Variables can be Chosen - Generalizes eitheror variables. Given a set of r binary variables (say x1,...,xr), at most k can be equal to one. The appropriate constraint is:
r

i=1

xi k

x1, x2, ..., xr = 0 or 1 If it is necessary that exactly k of the r projects must be chosen, this could be accomplished by incorporating the following constraint into the model:
r

i=1

xi = k

x1, x2, ..., xr = 0 or 1

Either-Or Constraints - Suppose there is a choice as to which of 2 resources to use for a certain purpose, so that it is necessary only that one of 2 constraints hold. Suppose that one of these constraints must hold: 3x1 + 2x2 18 or x1 + 4x2 16 Define M as an extremely large number and y as a binary variable (= 0 or 1). Then 3x1 + 2x2 18 + yM x1 + 4x2 16 + (1-y)M will guarantee that only one constraint is binding.

k out of r Constraints must Hold - Generalizes either-or constraints. Given a set of r constraints, only k of them must hold. The r constraints are f1 (x1,...,xn) b1 f2 (x1,...,xn) b2 . . fr (x1,...,xn) br Let yi = 0 or 1 for i = 1 to r. By the same logic as for either-or constraints, f1 (x1,...,xn) b1 + My1 f2 (x1,...,xn) b2 + My2 . . fr (x1,...,xn) br + Myr
r

yi = r - k
i=1

B. Goal Programming
- Goal Programming (GP): Branch of Linear Programming associated with solving multiobjective problems by minimizing the (weighted) sum of deviations from the goals/targets. Attributed to Charnes and Cooper [1950's]. Target Value - The desired value of attainment specified in a goal. Goal Deviation Variable - Often denoted Yi+ (over achievement of the target) and Yi- (underachievement of the target), these represent the solutions positive and negative deviations from the ith goal. Preemptive Priorities - A method of assigning priorities to the goals so that the relative satisfaction of a higher priority goal can not be sacrificed in the satisfaction of a lower priority goal.

Priority Weights - Assignment of weights to the goals so that the weighted sum of the deviations from the targets is minimized. Also called Penalty Points. Omnibus Objective Function - Mathematical statement that reflects all goals in a GP problem

Example - American Electronics, Inc.: American Electronics, Inc. produces color television sets on two production lines. The production rate of line 1 is two sets per hour, whereas it is 1.5 sets per hour in line 2. The regular production capacity is 40 hours a week for both lines. The expected profit from an average color television set is $100. The top management of the firm has the following goals for the week in ordinal ranking: G1 Meet the production goal of 180 sets for the week. G2 Limit the overtime of line 1 to ten hours. G3 Avoid the underutilization of regular working hours for both lines (assign differential weights according to the production rate of each line). G4 Limit the sum of overtime operation for both lines. Assign weights according to the relative cost of overtime hour. Formulate the above problem as a GP model.

Define: Xj = # production hours spent in line j per week, j = 1,2 Goal Deviation Constraints: Production (G1) 1.5x1 + 2.0x2 - (Y1+ - Y1-) = 180 (Y2+ - Y2-) = 40 Line 1 OT (G4) x1 + Line 1 OT > 10 hrs (G2) Y2 - (Y3+ - Y3-) = 10 Line 2 OT (G4) x2 - (Y4+ - Y4-) = 40 Combined Line Limit (G3) Y2- + Y4- - (Y5+- Y5-) = 0 Note that Goal 3 is be dealt with by using deviation variables from Goals 2 and 4. Now the omnibus objective function is Minimize Z = Y1- + Y3+ + Y5- + (Y2+ + Y4+)

Now we must still add our goal constraints 1.5x1 + 2.0x2 - (Y1+ - Y1-) = 180 (Production Target Constraint) x1 (Y2+- Y2-) = 40 (Line 1 OT Limitation) Y2+ - (Y3+ - Y3-) = 10 (Line 1 OT > 10 hrs) x2 - (Y4+- Y4-) = 40 (Line 2 OT Limitation) Y2 + Y4- - (Y5+- Y5-) = 0 (Combined Line Limit) + + + x1, x2, Y1 , Y1 , Y2 , Y2 , Y3 , Y3 , Y4+, Y4-, Y5+, Y5- 0 (Nonnegativity)

The optimal solution is x1 = Y1+ = Y2+ = Y3+ = Y4+ = Y5+ = 40, x2 = 0, Y1- = 0, Y2- = 0, Y3- = 20, Y4- = 0, Y5- = 60 0 0 10 0 0

In other words, produce 40 hours worth (60 units) on line 1 and 60 hours worth (120) units on Line 2 (with no deviation from the target production) with no Line 1 Overtime or Underutilization (and so 10 units under the maximum line 1 Overtime), no Line 2 Underutilization, and 20 hours of Line 2 Overtime.

If we are also told that we only have 190 picture tubes available and 951 legs (it takes four legs to build our televisions), we would also have the following constraints: (Picture Tube constraint) 1.5x1 + 2.0x2 190 1.5x1 + 2.0x2 951/4 (Leg constraint) Note that these constraints are often referred to as Hard Constraints while Goal Constraints are commonly called Soft Constraints.

The resulting formulation is Subject To: 1.5x1 + 2.0x2 - (Y1+ - Y1-) = 180 (Production Target Constraint) (Y2+- Y2-) = 40 (Line 1 OT Limitation) x1 Y2+ - (Y3+ - Y3-) = 10 (Line 1 OT > 10 hrs) x2 - (Y4+- Y4-) = 40 (Line 2 OT Limitation) Y2 + Y4- - (Y5+- Y5-) = 0 (Combined Line Limit) x1 + x2 190 (Picture Tube constraint) x1 + x2 237.75 (Leg constraint) + + x1, x2, Y1 , Y1 , Y2 , Y2 , Y3+, Y3-, Y4+, Y4-, Y5+, Y5- 0 (Nonnegativity) Minimize Z = Y1- + Y3+ + Y5- + (Y2+ + Y4+)

The new optimal solution is x1 = Y1+ = Y2+ = Y3+ = Y4+ = Y5+ = 40, x2 = 0, Y1- = 0, Y2- = 0, Y3- = 20, Y4- = 0, Y5- = 60 0 0 10 0 0

which is the exact same solution we achieved when we used only the Soft Constraints (why?).

Note that in this problem all goals are equally weighted. If we thought the fourth goal was six times as important as the other three goals, we could reformulate this problem as: Minimize Z = Y1- + Y3+ + Y5- + 6(Y2+ + Y4+) Subject To: x1 + x2 - (Y1+ - Y1-) = 180 (Production Target Constraint) .5x1 (Y2+- Y2-) = 40 (Line 1 OT Limitation) Y2+ - (Y3+ - Y3-) = 10 (Line 1 OT > 10 hrs) .67x2 - (Y4+- Y4-) = 40 (Line 2 OT Limitation) Y2++ Y3+ - (Y5+- Y5-) = 0 (Combined Line Limit) x2 190 (Picture Tube constraint) x1 + x1 + x2 237.75 (Leg constraint) + + x1, x2, Y1 , Y1 , Y2 , Y2 , Y3+, Y3-, Y4+ 0 (Nonnegativity) In both of these formulations we are utilizing Priority Weights (i.e., Penalty Points).

The new optimal solution is x1 = Y1+ = Y2+ = Y3+ = Y4+ = Y5+ = 40, x2 = 0, Y1- = 0, Y2- = 0, Y3- = 0, Y4- = 0, Y5- = 40 40 0 10 0 0

which is the exact same solution we achieved when we used only the Soft Constraints (why?).

If we wish to use a Preemptive Priority approach, we would: rate the goals in order of importance solve the problem minimizing deviation from the first target value (ignoring all other goals) add a constraint to the formulation forcing the first constraint to be satisfied at the level achieved in the previous step and solve the problem minimizing deviation from the second target value (ignoring all other goals) Continue this process until each goal has been addressed.

C. Negative Decision Variables

- Decision Variables that can assume negative values can easily be handled in a manner similar to deviation variables. Say that decision variable xi can assume either a positive or negative value. Then: Define xi = (xi+ - xi -), Substitute (xi+ - xi -) for xi everywhere in the formulation, and Make both xi+ and xi - nonnegative Note that xi+or xi - (or both) must equal zero xi+ is the amount that xi exceeds 0 xi- is the amount that xi is below 0

D. Nonlinear Programming
- Nonlinear Programming (NLP) - An approach to solving problems for which the objective function and/or at least one constraint has a term that involves i) an exponent that is a function of a decision variable and/or the product of multiple decision variables.