Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Introduction To Linear Programming: Algorithmic and Geometric Foundations of Optimization

Download as pdf or txt
Download as pdf or txt
You are on page 1of 28

Introduction to

Linear Programming

Algorithmic and Geometric Foundations of Optimization


Optimization and Linear Programming

• “Mathematical programming” is a class of methods for


solving problems which ask you to optimize: to maximize,
minimize, to find the best, the worst, the most, the least,
the fastest, the shortest, etc. Linear programming is the
simplest form, suitable only for problems with linear
objective functions and constraints.

•Two of the Basic Linear Programming Problems:

-The Product Mix Problem


•What quantities of products should be produced to maximize profit?
•In economics, this is the optimal allocation of scarce resources.

-The Blending Problem


•What combination of ingredients minimizes cost?
•For example, creating industrial chemical compounds.
Some Optimization Basics

•Unconstrained: max f äã x ëí = x
x

Easy: ∞

•Unconstrained: max f äã x ëí = x äã 1 − x ëí
x

df äã x ëí í 1
Still easy: = 1 − 2x = 0 àx =
dx 2

•In general, what are we doing?

-Set the derivative to zero and solve.


-Why?
•Necessary and Sufficient Conditions for an optimum over
single-peaked functions:

Necessity Sufficiency
df äã x ëí
=0 d 2 f äã x ëí
Maximum <0
dx dx
df äã x ëí
=0 d 2 f äã x ëí
Minimum >0
dx dx

f '(x) = 0 f "(x) < 0

f "(x) > 0
f '(x) = 0
Some Complications

•What about more complicated functions?

max f äã x, y, z ëí = x äã 1 − y ëí + y äã 1 − x ëí + z äã 1 − x − y ëí
x, y, z

•Calculus still works, but is cumbersome to do by hand. We


turn to computers and automate the process instead.

•What about constraints?


1
max f x = x 1 − x , subject to x #
ä
ã
ë
í
ä
ã
ë
í
x 3

•Again, calculus works (use the Lagrangian) for small


problems, but needs automation for larger problems.
More Complications…
Multiple Optima Discontinuous Functions
f(x) f(x)

x x
Uncertainty Integer Restrictions
f(x) f(x)

x x
The Need for Algorithmic Methods

•How can we automate searching for optima?

•Linear programming (LP) is one such method.

•LP is easy and fast, but it has some important limitations:

Limitations/Requirements Alternatives
Sometimes expectations can be used, but
(1) Decisions are made under certainty (1) generally stochastic programming is required
NLP (tough!). No guaranteed global optimizer
Objective function and constraints must be
(2) linear (2) other than enumeration! GA, GP, SA, heuristic
search, tabu, adaptive grid refinement, active sets
(3) Additivity (3) NLP

(4) Complete Independence (4) NLP

(5) Infinite Divisibility (5) Integer programming, branch & bound


LP Structure

•Assuming the requirements for LP have been met, what does a


LP look like?

•Decision Variables
-Quantities to be allocated
-Number of units to be produced
-Time intervals, unit quantities, proportions...

•Objective Function
-What is being maximized or minimized?
-Thought: How do we reduce the world to a set of
linear equations?

•Constraints
-(In)equalities specifying limits on the unbounded
optimization of a function.
-If Π = P H Q, why not just set Q to ∞?
Setting up a LP Problem: Product Mix

Fischbeck Electronics makes two models of television sets, A


and B. Fischbeck Electronics’ profit, ∏A, on A is $300 per set
and its profit on B, ∏B, is $250 per set. Thus, the problem facing
Fischbeck Electronics is:

max Π
¯ = ΠA x A + ΠB x B
xA , xB

Notice that unconstrained optimization would cause them to


produce only A! But, Fischbeck Electronics faces several
constraints:

(1) A Labor Constraint: They can’t use more than 40 hours of labor
per day (5 employees working regular shifts).
(2) A Manufacturing Constraint: They only have 45 hours of
machine time available (9 machines available for 5 hours each).
(3) A Marketing Constraint: They can’t sell more than 12 units of
set A per day.
Identifying the Problem Variables

•Decision Variables

-What can we vary?

x A and x B are the only variables we have direct control over.

•The Objective Function

-What are we maximizing (minimizing)?

Profit! (Cost!)
Thought: Max Profit = Min Cost = Max –Cost = ...

-How is it determined?
Π
¯ = Π A x A + Π B x B , but w e know that Π A = 300 and Π B = 250 .
So, the objective function is Π ¯ = 300x A + 250x B .
Identifying the Constraints

•3 primary constraints plus feasibility (x i $ 0 ú i ):

(1) Labor: A needs 2 hours of labor. B needs only 1 hour.


Total labor is equal to the number of units
produced times the labor required per unit.
2x A + x B # 40

(2) Manufacturing: A needs 1 hour of machine time. B


needs 3 hours.

x A + 3x B # 45

(3) Marketing: Can’t see more than 12 units of A per day.

x A # 12
The Entire LP and Possible Implications

max Π
¯ = 300x A + 250x B
xA , xB

such that: 2x A + x B # 40
x A + 3x B # 45
x A # 12
xA $ 0
xB $ 0

•What if x A or x B = 7.4235? What does it mean to


produce 0.4235 TV sets?

•Can 0.6689 people be employed?


Solving Linear Programming Problems

•Often, setting up LPs is the hardest part. This is because


the actual “solving” is usually done by computer.

•How do computers do it?

-We will use a graphical technique to illustrate the


intuition behind the algorithmic methods used in
optimization software packages.

•For any LP, there exists an infinite number of possible


solutions.

•Problem spaces may be vast. Many models have tens of


thousands of variables and constraints.

•How can we search? How can we limit the number of


possibilities we have to try?
Algorithmic Solutions

•The Simplex Algorithm (and variants)

-Created by George Dantzig in 1947 for the Air Force


Office of Scientific Research
-“If the LP problem has an optimal solution, it will be
found in a finite number of iterations.” (Note: this isn’t
always very comforting)
-The simplex algorithm finds the solution by making
iterated improvements from an initial position (solution).

•The Ellipsoid Algorithms

-The best known example is Karmarkar’s Algorithm


(created at AT&T Bell Labs)
-Runs in linear time
-Faster than simplex for very large problems
-We won’t worry about Karmarkar...
What about the Solver?
•Do not use Excel or Solver as a black box! Know what’s
going on underneath the surface. Remember: you will always
get an answer back from Solver. However, knowing whether
or not that answer is relevant requires understanding what
Solver is doing. Know its limitations!

•Linear programs: Standard simplex

•Quadratic programs: Modified simplex (QP is technically


NLP, but results specific to quadratic forms can be exploited to
allow for significant simplification.

•Nonlinear Programs: Generalized Reduced Gradient


Method. NOTE: Unless you check Assume linear model, it
will automatically use GRG2. GRG2 may have difficulty with
some LP or QP problems that could have more easily been
solved by another method. Beware the black box!

•Integer programs: Branch and Bound


Solution Method Intuition

•All of the solution methods just mentioned tend to have


very intuitive explanations.

•We will talk about how the simplex method works by


graphically solving a small (two variable) linear program.

•Again, our problem:

max Π
¯ = 300x A + 250x B
xA , xB

such that: 2x A + x B # 40
x A + 3x B # 45
x A # 12
xA $ 0
xB $ 0
Graphical Solutions

•Limited to either 2 constraints/variables.

•The objective function: max Π


¯ = 300x A + 250x B
xA , xB

•The solution is contained somewhere in real-valued


space. That’s a big area. Fortunately, we can narrow it
down immediately to the first quadrant because of the
feasibility constraints: x A $ 0 and x B $ 0 .

x(B)
1001010101 1
0
1100010101010101001
1
10010101010101010101
1
0 1
0 1
0

x(A)
1
0
Step One: Graphing the Constraints

•Consider Constraint #1: 2x A + x B # 40

•When x A = 0, x B # 40. To anchor the line, note that when


x B = 0, x A # 20.

x(B)

2x(A) + x(B) = 40
1001010101010101010101010101010101010101
0101010101010101010101010101010101010101
0101010101010101010101010101010101010101
40

2x(A) + x(B) < 40

x(A)

20
Constraint #2: x A + 3x B # 45
•When x A = 0, then x B # 15. Anchoring the line, when
x B = 0, then x A # 45.

x(B)

x(A) + 3x(B) = 45
10010101010101010101
0101010101
01010101010101010101
0101010101
01010101010101010101
0101010101
01010101010101010101
15 x(A) + 3x(B) < 45

x(A)
45
1 1
0 1
0 1
0
x(B)

100101010101010101010101010101010 10
0 10101
1 1
0 1
0 1
0

12
010101010101010101010101010101010
10
0 10101
•This constraint requires no solving for.
Constraint #3: x A # 12

x(A)
15
40
x(B)

0
1 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 1
0 0
1 0
1
11000101010101010101010101010101010101010101010101010 1
1
0
0
1
1
0
1
0
1
0
1 10
0 1

12
01010101010101010101010101010101010101010101010101010 1
0
1
says nothing about the optimal set!

20
0101010101010101010101010101010
010101010101010101010101010101010101010 1
01010101010101010101
010101010101010101010101
010101010101
(Profit)

010101010101010101010101
Feasible Set

45
x(A)
010101010101010101010101
Combining the Constraints: The Feasible Set

solutions (if a solution exists). Keep in mind that this is the feasible set - it
•The intersection of all constraint-feasible sets represents the set of feasible
Identifying the Optimal Points Within the Feasible Set
•Method #1: Corner-Point Enumeration

(Close-Up)

(0,15)
15 (?,?)

(12,0)

(0,0) 0 12

•The “mystery” point is the intersection of the two lines.


Two equations, two unknowns: x A = 12 and x A + 3x B = 45.
Therefore 12 + 3x B = 45 à x B = 11.

•Thus, the coordinates of the fourth corner are (12, 11).


Choosing a Corner-Point
•Each of the corner-points represents a possible optimal
solution. To identify the optimal one, substitute them into
the objective function and see which one maximizes the
function.

Trial Objective Function Profit


1 300(0) + 250(0) $-0-
2 300(0) + 250(15) $3,750
3 300(12) + 250(0) $3,600
⁄4⁄ 300(12) + 250(11) $6,350

•But why limit ourselves to corner-points? Why shouldn’t


we choose a point in the interior of the feasible set?
Isoprofit (Isocost) Curves
•x A and x B can take many values. To identify the optimal values,
substitute arbitrary values in as the objective function and then
change them to push the frontier out towards the boundary of the
feasible set.

•To start, pick a value for the objective function - say, $1,500. This
becomes an isoprofit curve: 300x A + 250x B = 1500.

•When x A = 0, x B = 6 and when x B = 0, x A = 5 .

x(B)

300x(A) + 250x(B)=1500
1001010101010101
0101010101010101
01010101
0101010101010101
01010101
0101010101010101

x(A)
5
Adding Constraints to the Isoprofit Curves
x(B)

300(12) + 250(11) = 6,350

15
1
0
1
0
10
1
1
0
1
0
10
1
1
0
10
1
1001010101010101010101010 1
0101010101010 1
0
0101010101010101010101010 1
1
0
0101010101010
0101010101010101010101010
1
0
11

x(A)

12

•Any further advancement of the isoprofit curve would


cause it to leave the feasible set. Thus, (1) optimal
solutions exist only on boundaries, and (2) if only one
exists, the optimal solution must always occur at a vertex.
A Quick Review
•LPs are good for optimization problems involving
maximizing profits and minimizing costs.

•LP = Decision Variables, Objective Function, and Constraints

•Decision Variables are the quantities of resource being


allocated

•The Objective Function is what’s being optimized.

•Constraints are resource limitations or requirements

•Advantages of Linear Programming:

-Relatively quick
-Guaranteed to find optimal solution
-Provides natural sensitivity analysis (shadow prices)
Disadvantages of Linear Programming

•Absence of risk (does expectation adequately capture


risk?)

•Restriction to linear objective functions

-No correlation among variables


-No positive or negative synergies
-Nonlinear programming is very difficult

•Fractional solutions often have no meaning

•Reducing the world to a set of linear equations is usually


very difficult
Solution Methods

•The Simplex Method:

(1) Pick a trial solution point


(2) Move out from that point along the edges,
looking for vertices.
(3) Keep moving along the positive gradient until
no further improvements can be made.

•The Graphical Analogue to the Simplex:

(1) Draw the constraints


(2) Push the isoquant as far in the optimal
direction as possible.
(3) Pick the last vertex exiting the feasible set.

•For more graphing help, see The Geometry of Linear


Programming at http://equilibrium.hss.cmu.edu/dadss/

You might also like