Operations Research 1:

Reformulating “Non-standard” Models

M. Mujiya Ulkhaq, Ph.D.

Department of Industrial Engineering
Diponegoro University, Semarang, Indonesia For undergraduate (bachelor) level,
E-mail: [email protected] Department of Industrial Engineering,
Homepage: https://sites.google.com/view/mujiyaulkhaq/ Diponegoro University, Indonesia
 Introduction
• Recall our “standard form” of the linear programming problem has the following features:
o The objective function is to be maximized.
o The functional constraints are in ≤ form with non-negative right-hand sides.
o The variables have non-negativity constraints.
• Sometimes, we face a linear programming problem in which is not in our standard form, e.g.,
o having negative right-hand sides;
o variables allowed to be negative;
o having equality constraints;
o having functional constraints in ≥ form.
• This “non-standard form” linear programming model can be reformulated as a convenient artificial problem for
preparing to apply the simplex method using artificial variables.
• Two alternative methods are available for doing this:
• The big M method; and
• The two-phase method.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 2

 Negative Right-Hand Sides
• Simple rule: “If the right-hand side (RHS) is negative, multiply through both sides by −1 first.”
• Example:
x1 − x2 ≤ −1 is equal to −x1 + x2 ≥ 1
3x1 + 2x2 − x3 ≤ −6 is equal to −3x1 − 2x2 + x3 ≥ 6
• Using this procedure, now the right-hand side is positive.
• Having non-negative right-hand sides for all the functional constraints enables the simplex method to begin.
• However, we now have the “non-standard form” since the functional constraints are in ≥ form.
• The artificial-variable technique, such as the big M method and two-phase method can be used to solve this.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 3

 Variables Allowed to be Negative
Variables with a Bound on the Negative Values Allowed
• Consider any decision variable xj that is allowed to have negative values which satisfy a constraint of the form:
xj ≥ Lj, where Lj is some negative constant.
• This constraint can be converted to a non-negativity constraint by making the change of variables:
x′j = xj − Lj, so x′j ≥ 0.
• Thus, x′j + Lj would be substituted for xj throughout the model, so that the new decision variable x′j is positive.

Example:

x′1 = x1 + 10
x1 = x′1 – 10
M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 4
 Variables Allowed to be Negative (2)
Variables with No Bound on the Negative Values Allowed
• In the case where xj does not have a lower-bound constraint in the model formulated, another approach is:
xj is replaced throughout the model by the difference of two new nonnegative variables:
xj = xj+ − xj−, where xj+ ≥ 0, xj− ≥ 0.
• Since xj+ and xj− can have any non-negative values, this difference (xj+ − xj−) can have any value (positive or
negative).

Example:

x1 is unrestricted in sign
x1 = x1+ – x1−

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 5

 The Big M Method

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 6

 The Big M Method: An Introduction
• The Big M method is designed to force the simplex method to drive all the artificial variables to zero by imposing
a huge penalty on having values greater than zero.
• It does this by introducing a quantity denoted by M that symbolically represents a huge positive number that is
vastly larger than any of the actual numbers in the real problem.
• It is not necessary to assign it a specific value, but some people find it useful to think of M as a “million”.
• Assigning a penalty of M times an artificial variable for each of the artificial variables should enable the simplex
method to force the values of these variables down as low as possible.
• Since the artificial variables have nonnegativity constraints, forcing their values down should take them down to
zero.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 7

 Conceptual Outline of the Big M Method
To prepare for applying the Big M method, a linear programming model in non-standard form needs to be
reformulated as a convenient artificial problem.
1. This reformulation includes introducing artificial variable, which has the effect of revising the original
problem by enlarging the feasible region.
The purpose of creating the artificial problem is that it provides an initial BF solution for this problem to enable starting the
simplex method.
2. The objective function then needs to be revised by subtracting from the original objective function an
additional term for each artificial variable.
This additional term being subtracted is M times the artificial variable. Since the objective function is being maximized,
each such new term will enable the simplex method to drive that variable to zero in due time.
3. Before the simplex method can begin working on the artificial problem, the entire system of equations
(including Eq. (0)) needs to be in proper form from Gaussian elimination.
Algebraic operations need to be performed to algebraically eliminate these artificial variables from Eq. (0).
4. Apply the simplex method to the artificial problem until all of the artificial variables have been driven to values
of zero after some iterations by converting all these basic variables to non-basic variables.
This provides a BF solution for the real problem.
5. Now apply the simplex method to the real problem until it reaches an optimal solution.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 8

 The Big M Method: Equality Constraints
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
1. Revise the problem by introducing an artificial variable r1 that is inserted into the equality constraint:

Maximize Z = 3x1 + 5x2 Maximize Z = 3x1 + 5x2

Subject to x1 ≤ 4 Subject to x1 ≤ 4
2x2 ≤ 12 2x2 ≤ 12
3x1 + 2x2 = 18 3x1 + 2x2 + r1 = 18
x1 ≥ 0 and x2 ≥ 0 x1 ≥ 0, x2 ≥ 0, r1 ≥ 0

2. Revise the objective function by subtracting from the original objective function an additional term for this
artificial variable:
Maximize Z = 3x1 + 5x2 Maximize Z = 3x1 + 5x2 – Mr1
Subject to x1 ≤ 4 Subject to x1 ≤ 4
2x2 ≤ 12 2x2 ≤ 12
3x1 + 2x2 + r1 = 18 3x1 + 2x2 + r1 = 18
x1 ≥ 0, x2 ≥ 0, r1 ≥ 0 x1 ≥ 0, x2 ≥ 0, r1 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 9

 The Big M Method: Equality Constraints (2)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
3. Add slack variables:
(0) Z = 3x1 + 5x2 – Mr1 (0) Z – 3x1 – 5x2 + Mr1 =0
(1) x1 ≤ 4 (1) x1 + s1 =4 initial basic variables:
(2) 2x2 ≤ 12 (2) 2x2 + s2 = 12 (s1, s2, r1)
(3) 3x1 + 2x2 + r1 = 18 (3) 3x1 + 2x2 + r1 = 18

4. This system is not yet in proper form from Gaussian elimination because a basic variable r1 has a non-zero
coefficient in Eq. (0) (Recall that all basic variables must be algebraically eliminated from Eq. (0).)
Eliminate r1 from Eq. (0) algebraically: subtract Eq. (0) M times Eq. (3)
Z – 3x1 – 5x2 + Mr1 = 0
M3x1 + M2x2 + Mr1 = 18M
Z – (3M + 3)x1 – (2M + 5)x2 = –18M

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 10

 The Big M Method: Equality Constraints (3)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
5. Apply the simplex method:
Remember M
is a huge number!
so 3M+3 > 2M+5

Optimal Solution:
basic variables:
(x1, s1, x2) = (2, 2, 6)
Z = 36

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 11

 The Big M Method: Constraints in ≥ form
Consider the following problem:
Maximize Z = x1 + 5x2
Subject to 3x1 + 4x2 ≤ 6
x1 + 3x2 ≥ 2
x1 ≥ 0 and x2 ≥ 0

1. Introduce the surplus variable

Maximize Z = x1 + 5x2 Maximize Z = x1 + 5x2
Subject to 3x1 + 4x2 ≤ 6 Subject to 3x1 + 4x2 ≤ 6
x1 + 3x2 ≥ 2 x1 + 3x2 – s2 = 2
x1 ≥ 0 and x2 ≥ 0 x1 ≥ 0, x2 ≥ 0, s2 ≥ 0

s2 is called a surplus variable because it subtracts the surplus of the left-hand side over the right-hand side to
convert the inequality constraint (≥) to an equivalent equality (=) constraint.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 12

 The Big M Method: Constraints in ≥ form (2)
2. Add the artificial variable
Maximize Z = x1 + 5x2 Maximize Z = x1 + 5x2
Subject to 3x1 + 4x2 ≤ 6 Subject to 3x1 + 4x2 ≤ 6
x1 + 3x2 – s2 = 2 x1 + 3x2 – s2 + r1 = 2
x1 ≥ 0, x2 ≥ 0, s2 ≥ 0 x1 ≥ 0, x2 ≥ 0, s2 ≥ 0, r1 ≥ 0
3. Revise the objective function
Maximize Z = x1 + 5x2 Maximize Z = x1 + 5x2 – Mr1
Subject to 3x1 + 4x2 ≤ 6 Subject to 3x1 + 4x2 ≤ 6
x1 + 3x2 – s2 + r1 = 2 x1 + 3x2 – s2 + r1 = 2
x1 ≥ 0, x2 ≥ 0, s2 ≥ 0, r1 ≥ 0 x1 ≥ 0, x2 ≥ 0, s2 ≥ 0, r1 ≥ 0
4. Add the slack variable
Maximize Z = x1 + 5x2 – Mr1 Maximize Z = x1 + 5x2 – Mr1
Subject to 3x1 + 4x2 ≤ 6 Subject to 3x1 + 4x2 + s1 = 6
x1 + 3x2 – s2 + r1 = 2 x1 + 3x2 – s2 + r1 = 2
x1 ≥ 0, x2 ≥ 0, s2 ≥ 0, r1 ≥ 0 x1 ≥ 0, x2 ≥ 0, s1 ≥ 0, s2 ≥ 0, r1 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 13

 The Big M Method: Constraints in ≥ form (3)
5. Eliminate basic variable from Eq. (0)
Eliminate r1 from Eq. (0) algebraically: subtract Eq. (0) M times Eq. (2)
Z – x1 – 5x2 + Mr1 = 0
Mx1 + 3Mx2 – Ms2 + Mr1 = 2M
Z – (1 + M)x1 – (5 + 3M)x2 + Ms2 = –2M

Final (with surplus, slack, and artificial variable) problem formulation:

Maximize Z – (1 + M)x1 – (5 + 3M)x2 + Ms2 = –2M

Subject to 3x1 + 4x2 + s1 = 6
x1 + 3x2 – s2 + r1 = 2
x1 ≥ 0, x2 ≥ 0, s2 ≥ 0, r1 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 14

 The Big M Method: Constraints in ≥ form (4)
6. Apply the simplex method:

Optimal Solution:
basic variables:
(s2, x2) = (5/2, 3/2)
Z = 15/2

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 15

 The Two-Phase Method

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 16

 Two-Phase Method: An Introduction
• Another alternative to solve the “non-standard form” of linear programming models is the two-phase method.
• This method enables applying the simplex method directly without introducing the huge number M.
• For the previous Wyndor Glass Co. problem (with the equality constraint), the objective function is:
o Real problem Maximize Z = 3x1 + 5x2
o Big M method Maximize Z = 3x1 + 5x2 – Mr1
o Two-phase method Phase 1 : Maximize Z = – r1 (until r1 = 0 or as non-basic variable)
Phase 2 : Maximize Z = 3x1 + 5x2 (with r1 = 0)

• The Phase 1’s objective function is obtained by dividing the Big M method’s objective function by M and then
dropping the negligible terms (i.e., the decision variables).
• Since Phase 1 concludes by obtaining a BF solution for the real problem (one where r1 = 0), this solution is then
used as the initial BF solution for applying the simplex method to the real problem (with its real objective
function) in Phase 2.

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 17

 Summary of the Two-Phase Method
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 1: The objective of Phase 1 is to find a BF solution for the real problem.
o Reformulate the “non-standard form” by introducing artificial and/or surplus variables as needed.

Phase 1 : Maximize Z = – ∑$!"# 𝑟! , where ri is the artificial variable (i = 1, 2, …, L).

Subject to “reformulated” constraints

For the Wyndor Glass Co. problem:

Phase 1 : Maximize Z = –r1
Subject to x1 + s1 = 4
2x2 + s2 = 12
3x1 + 2x2 + r1 = 18
x1 ≥ 0, x2 ≥ 0, s1 ≥ 0, s2 ≥ 0, r1 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 18

 Summary of the Two-Phase Method (2)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 1: The objective of Phase 1 is to find a BF solution for the real problem.
o Eliminate the basic variables from Eq. (0)
Eliminate r1 from Eq. (0): subtract Eq. (3) from Eq. (0)
Z + r1 = 0
3x1 + 2x2 + r1 = 18
Z – 3x1 – 2x2 = –18

Final problem formulation of Phase 1:

Maximize Z – 3x1 – 2x2 = –18
Subject to x1 + s1 = 4
2x2 + s2 = 12
3x1 + 2x2 + r1 = 18
x1 ≥ 0, x2 ≥ 0, s1 ≥ 0, s2 ≥ 0, r1 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 19

 Summary of the Two-Phase Method (3)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 1: The objective of Phase 1 is to find a BF solution for the real problem.
o Apply the simplex method

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 20

 Summary of the Two-Phase Method (4)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 2: The objective of Phase 2 is to find an optimal solution for the real problem.
Starting from the BF solution obtained at Phase 1, use the simplex method to solve the real problem.

For the Wyndor Glass Co. problem:

Phase 2 : Maximize Z = 3x1 + 5x2
Subject to x1 + s1 = 4
2x2 + s2 = 12
3x1 + 2x2 = 18
x1 ≥ 0, x2 ≥ 0, s1 ≥ 0, s2 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 21

 Summary of the Two-Phase Method (5)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 2: The objective of Phase 2 is to find an optimal solution for the real problem.
o Drop the artificial variables since they are not part of the real problem.
o Substitute Phase 2’s objective function into Eq. (0).
artificial variable r1 is dropped
From the BF solution obtained at Phase 1:

Eq. (1) x1 – 1/3 s2 = 2 ⬄ x1 = 2 + 1/3 s2

Eq. (3) x2 + 0.5 s2 = 6 ⬄ x2 = 6 – 0.5 s2

Phase 2’s objective function : Maximize Z = 3x1 + 5x2

Maximize Z = 3(2 + 1/3 s2) + 5(6 – 0.5 s2)
Maximize Z = 36 – 1.5 s2

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 22

 Summary of the Two-Phase Method (6)
Consider the Wyndor Glass Co. problem where the functional constraint 3x1 + 2x2 ≤ 18 has been changed to 3x1 + 2x2 = 18.
• Phase 2: The objective of Phase 2 is to find an optimal solution for the real problem.
o Eliminate the basic variables from Eq. (0) – if any
o Apply the simplex method – if necessary

Optimal Solution:
basic variables: (x1, s1, x2) = (2, 2, 6)
Z = 36

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 23

 The Two-Phase Method: Example
Consider the following problem:
Maximize Z = 3x1 – x2 + 2x3
Subject to x1 + 3x2 + x3 ≤ 5
2x1 – x2 + x3 ≥ 2
4x1 + 3x2 – 2x3 = 5
x1 ≥ 0; x2 ≥ 0; x3 ≥ 0

• Phase 1 : Maximize Z = – r1 – r2
Subject to x1 + 3x2 + x3 + s1 = 5 s1 is the slack variable
2x1 – x2 + x3 – s2 + r1 = 2 s2 is the surplus variable
4x1 + 3x2 – 2x3 + r2 = 5 r1 and r2 are the artificial variables
x1, x2, x3, s1, s2, r1, r2 ≥ 0

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 24

 The Two-Phase Method: Example (2)
• Phase 1
o Eliminate the basic variables from Eq. (0)
Eliminate r1 and r2 from Eq. (0): subtract Eq. (2) and (3) from Eq. (0)
Eq. (0) Z + r1 + r2 = 0
Eq. (2) 2x1 – x2 + x3 – s2 + r1 =2
Eq. (0)* Z – 2x1 + x2 – x3 + s2 + r2 = –2
Eq. (3) 4x1 + 3x2 – 2x3 + r2 = 5
Eq. (0)** Z – 6x1 – 2x2 + x3 + s2 = –7

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 25

 The Two-Phase Method: Example (3)
• Phase 1
o Apply the simplex method

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 26

 The Two-Phase Method: Example (4)
• Phase 2
o Drop the artificial variables.
o Substitute Phase 2’s objective function into Eq. (0). artificial variables r1 and r2
are dropped
From the BF solution obtained at Phase 1:

Eq. (1) x1 + 0.1 x3 – 0.3 s2 = 1.1 ⬄ x1 = 1.1 – 0.1 x3 + 0.3 s2

Eq. (3) x2 – 0.8 x3 + 0.4 s2 = 0.2 ⬄ x2 = 0.2 + 0.8 x3 – 0.4 s2

Phase 2’s objective function : Maximize Z = 3x1 – x2 + 2x3

Maximize Z = 3(1.1 – 0.1 x3 + 0.3 s2) – (0.2 + 0.8 x3 – 0.4 s2) + 2x3
Maximize Z = 2.9 + 0.9 x3 + 1.3 s2

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 27

 The Two-Phase Method: Example (5)
• Phase 2
o Eliminate the basic variables from Eq. (0) – if any
o Apply the simplex method – if necessary

Optimal Solution:
basic variables: (x3, x1, s2) = (2.5, 2.5, 5,5)
Z = 12.3

M. Mujiya Ulkhaq, Ph.D. Reformulating “non-standard” models 28