Slajd 1

Linear Programming

• A linear programming (LP) problem may be

defined as the problem of maximizing or
minimizing a linear function subject to linear
constraints, the constraints may be equalities or
inequalities

• Many optimization problems, especially in

economics, can be formulated as linear programs

• The most popular solution method of linear

programming is Simplex

• The author of Simplex – George Dantzig –

obtained in 1975 the Nobel prize in economy

Formulations (1)

x = [x

, x

,…, x

]

column vector of variables (n

elements)

b = [b

, b

,…, b

]

column vector of constants

(m elements)

c = [c

, c

,…, c

] row vector of constants (n

elements)

matrix of constants (m x n

elements)





...

Formulations (3)

indices
j = 1,2,…,n

variables

i = 1,2,…,mconstraints
constants
a

coefficient of variable j in constraint i

right-hand side of constraint i

cost of variable j

Formulations (4)

variables
x

variable j

objective
min z = 

constraints


= b

for

i = 1,2,…,m

 0

for

j = 1,2,…,n

Formulations (5)

Inequality constraints of type 



 b

for i = 1,2,…,m

can be writen as equality constraints using non-

negative slack variables in the following way



+ z

= b

, z

 0 for i = 1,2,…,m

Inequality constraints of type 



 b

for i = 1,2,…,m

can be writen as equality constraints using non-

negative slack variables in the following way



– z

= b

, z

 0 dla i = 1,2,…,m

Example (1)

max 2x

+ x

s.t.
x

+ x

 6

-2x

+ x

 3

 0

solution
x

= 6 x

= 0

Example (2)

max x

+ 2x

s.t.
x

+ x

 6

-2x

+ x

 3

 0

rozwiązanie
x

= 1 x

= 5

Simplex Algorithm

• Simplex method was developed in 1947 by Georga

Dantziga

• The main idea is the search along vertices and

edges of the solution space polytype

• The algorithm moves to vertices with higher

objective function value

• If a local maximum is reached, by convexity it is also

the global optimum and the algorithm terminates

• It also can terminate when an unbounded edge is

visited, concluding that the problem has no solution

• The algorithm always terminates since the number of

vertices in the polytope is finite

Duality of Linear Programming

(1)

• Many optimization algorithms applies the duality

of Linear Programming

• For a LP problem called the primal problem, a

dual problem can be formulated

• Correspondence between properties of the

primal problem and the dual problem are used in
the construct of optimization methods in order to
reduce the solution time

Duality of Linear Programming

(2)

Primal problem
variables
x

variable j

objective
min 

constraints


 b

for i = 1,2,

…,m

 0 for j = 1,2,…,n

Dual problem
variables



dual variable i

objective
max 



constraints




 c

for j = 1,2,…,n



 0 for i = 1,2,…,m

Duality of Linear Programming

(4)

Theorem. The dual problem of the dual problem is

the primal problem

Theorem. If x is a feasible solution of the primal

problem and  is a feasible solution of the dual

problem, then the value of the primal problem
objective function cannot be lower than the value
of the dual problem objective function

cx  b

Duality of Linear Programming

(5)

Theorem. (strong duality). If for a pair of feasible

solutions (x

,

) hold the following condition cx

b

then (x

,

) is a pair of optimal solutions

Theorem. Either both problems (primal and dual)

are solvable simultaneously, or one of the two is
degenerate in the sense that it does not admit
feasible vectors

Integer Problems

• Problems with some continous variables and

some integer variables are called Mixed Integer
Programming (MIP)

• Problems with all integer variables are called

Integer Programming (IP)

• IP and MIP problems are mostly NP-hard
• A special class of integer problems are problems

with binary variables

• The basic method to solve integer problems is

branch and bound approach

MIP Formulation (1)

indices
j = 1,2,…,n

variables

i = 1,2,…,mconstraints
constants
a

coefficient of variable j in constraint i

right-hand side of constraint i

cost of variable j

MIP Formulation (2)

variables
x

variable j

objective
min z = 

constraints


= b

for

i = 1,2,…,m

 {0,1} for

j = 1,2,…,k

 0

for

j = k + 1, k + 2,…,n

IP Formulation

variables
x

variable j

objective
min z = 

constraints


= b

for

i = 1,2,…,m

 {0,1,...,M}

for

j = 1,2,…,n

Binary MIP Relaxation

Let N

and N

denote sets of of variables with set

values 0 and 1, respectively. Let N

be a set of

variables with not fixed values

A MIP problem can be relaxed to a LP problem as

follows

continuous

for

j  N

0  x

 1 for

j  N

= 0

for

j  N

= 1

for

j  N

A solution of the above problem (if exists) is a lower

bound of the orignal problem

Example (1)

min z = -6x

– 5x

s.t. 3x

+ x

 11

-x

+ 2x

 5

, x

 0

LP solution
(not integer)
x

= 2

= 3

z = -33

Example (2)

min z = -6x

– 5x

s.t. 3x

+ x

 11

-x

+ 2x

 5

, x

 0,

integer

solution
x1 = 3 x2 = 2
z = -28

Branch and Bound (2)

• Bounding enables to verify if the current solution

can yield objective value better than the current best
value

• To find the lower (upper) bound usually a relaxed

version of the problem is solved

• Branching is applied to select a variable that is

used to split the problem into two smalleer
subproblems

• The efficiency of the method depends strongly on

the branching and bounding functions

• The subsquent steps of the algorithm can be

presented in a form of solution tree

Branch and Bound (3)

procedure BBB (N

)

begin

solution (N

, x, z) {solving the relaxed linear poblem};

if N

= ∅ or for all i  N

are binary then

if z < z

best

then begin z

best

:= z; x

best

:= x end

else {in the current solution the value of x

is not binary for some i  N

}

if z ≥ z

best

then

return {bounding}

else

begin {branching}

choose i  N

such that x

is fractional;

BBB (N

\{i},N

 {i},N

);

BBB (N

\{i},N0,N

 {i})

end

end { procedure }

Branch and Cut

• Branch and cut is an improved branch and bound

method

• The effectivness of the branch and cut method

depends strongly on the bounds

• Additional constraints called cuts or valid

inequalities are introduced to the problem

• Cut inequalities enable calculation of more effective

bounds

• Global cuts are valid for all feasible integer solutions

and are added to the problem formulation

• Local cuts are valid only for the currently considered

branch and bound subtree

Example (1)

min z = -6x

– 5x

s.t. 3x

+ x

 11

-x

+ 2x

 5

, x

 0,

integer

LP solution
(not integer)
x

= 2

= 3

z = -33

Example (3)

min z = -6x

– 5x

s.t.

+ x

 11

-x

+ 2x

 5

 3

, x

 0,

integer
LP solution (integer)
x

= 3 x

= 2 z = -28

Completed
subproblem

Example (4)

min z = -6x

– 5x

s.t. 3x

+ x

 11

-x

+ 2x

 5

 2

, x

 0, integer

LP solution (not integer)
x

= 2 x

= 3,5 z =

-29,5
Since the solution is
not integer, the
algorithm continues

Example (5)

New cut is introduced
min z = -6x

– 5x

s.t.

+ x

 11

-x

+ 2x

 5

 2

+ x

 7 (cut)

, x

 0, integer

LP solution (not integer)
x

= 1,8

= 3,4 z =

-27,8

The objective is larger

than the best

solution. The

algorithm stops

Nonlinear Programming

• A nonlinear programming (NLP) problem may

be defined as the problem of maximizing or
minimizing a function subject to some
constraints, at least the objective or one of the
constraints is not linear

• If the objective function and constraints have

special properties (e.g., there are convex)
optimality criteria for the problem can be
formulated

• The linearization of the nonlinear functions can

be applied as an approximation solution method

Formulations (1)

NLP with inequality constraints
min f(x)

xR

s.t.
g

(x)  0

for i  I

NLP with equality constraints
min f(x)

xR

s.t.
h

(x) = 0

for j  E

NLP Optimality (1)

Definition. All points that statisfy constraints of the

NLP are called feasible points

Definition. Direction d from point x satisfying all

constraints is feasible, if there exists



> 0 that

for any 0 





the point x +



d is feasible

(holds all constraints of the NLP)

Definition. In a feasible point x constraints of the

NLP can be

• Active included in a set A(x) = {i  I : g

(x) = 0}

• Nonactive, for which g

(x) < 0

NLP Optimality (2)

Let functions f : X = R

 R

and g

: X = R

 R

differentiable and let x



be a local minimum of the

NLP

Let A(x



) be a set of active constraints in x



The feasible direction d in x



must satisfy





d)  0 for each i  A(x



)

Using Taylor's theorem we can write





d) = g



) +



g



)

d + O(



)  0

Thus, the sufficient condition for d to be feasible is

g



)

d  0 for each i  A(x



)

NLP Optimality (3)

Directions satysfying the above condition are

feasible only when some additional conditions
called regularity conditions holds

If point x



is a minimum, then on each feasible

direction the objective function must increase

Therefore, in optimal point x



for each feasible

direction d the following condition must be
satisfied

f(x



)

d  0

NLP Optimality (4)

Farkas' lemma. Conditions g



)

d  0 for each i

 A(x



) and f(x



)

d  0 means that the objective

function derivative vector f(x



) is in a convex

cone defined by vectors {-g



), i  A(x



)}

In other words, the vector of the objective function

in the optimum point can ge writen as {-g



), i

 A(x



)} of derivatives in this point

f(x



) =  

iA(x)



g



)

Coefficients



are called Lagrange multipliers

For each nonactive constraint, multipliers are 0

Kuhn-Tucker Conditions

• The Kuhn-Tucker theorem is a theorem in

nonlinear programming

• Kuhn-Tucker theorem states that if a regularity

condition holds and the objective function f and
the functions g

and h

are convex, then a solution

x which satisfies the conditions g

and h

for a

vector of multipliers lambda is a global minimum

• The Kuhn-Tucker theorem is a generalization of

Lagrange multipliers

• Farkas's lemma is key in proving this theorem

Regularity conditions

• All constraints g

are linear functions

• All constraints g

are convex functions and there

is a point x, that g

(x) < 0 holds for all i  I

• The gradients of the active inequality constraints

and the gradients of the equality constraints are
linearly independent

Formulations (1)

Necessary Kuhn-Tucker conditions for a problem

with inequality constraints

1. f(x) + 

iI



g

(x) = 0

2. g

(x)  0 for i  I



 0

for i  I



(x) = 0

for i  I

or 

iI



(x) = 0

In order for a solution point to satisfy the above KKT

conditions, it should satisfy some regularity
condition

Formulations (2)

Necessary Kuhn-Tucker conditions for a problem

with equality constraints

1. f(x) + 

jE



h

(x) = 0

2. h

(x) = 0 for j  E

In order for a solution point to satisfy the above KKT

conditions, it should satisfy some regularity
condition

Formulations (3)

Necessary Kuhn-Tucker conditions for a problem

with equality and inequality constraints

1. f(x) + 

jE



h

(x) + 

iI



g

(x) = 0

2. g

(x)  0 for i  I

3. h

(x) = 0 for j  E



 0

for i  I



(x) = 0

for i  I

In order for a solution point to satisfy the above KKT

conditions, it should satisfy some regularity
condition

Example (1)

indices
i = 1,2,…,n

links (variables)

constants
c

capacity of link i

suma of all link flows

variables
f

flow in link i

Example (2)

objective
minimize T = 

/(c

– f

)

constraints
f

 c

for i = 1,2,…,n

(constraint g

)

D – 

= 0

(constraint g)

Example (3)

Each constraint is linear, the objective funtion is

convex, thus Kuhn-Tucker conditions are sufficient
and and the global minimum can be found

KT constraints can be writen as follows
c

/(f

– c

)

–



= 0 for i = 1,2,…,n (objective)

– c

)



= 0

for i = 1,2,…,n

(constraint g

)

(D – 

)



= 0

(constraint g)



 0

for i = 1,2,…,n

Example (4)

Using the first constraint we obtain

If we omit „+” we have, f

 c

, what means that

constraint g

is satsified for i = 1,2,…,n

Due to the construct of the objective function we

must have
f

< c

, what yields the following condition



= 0 dla i = 1,2,…,n

Consequently, we can write











Example (5)

Using this formula in constraint g we have

Next, using the above formula we obtain

The solution holds all Kuhna-Tucker constraints,

thus it is optimal





















i i











i i

Lagrangian Function (1)

Lagrangian function is defined in the following

way

L(x,) = f(x) + 

jE



(x) + 

iI



(x) = 0

where  is a vector of multipliers,   R

Definition. Point <x



,



>, 



 0, 



 R

, x



 X is a

saddle point of L(x,) function if,

L(x



,



)  L(x,



), x  X

L(x



,



)  L(x



,),   0,   R

Lagrangian Function (2)

Theorem. Let   0 and x

 X. Point <x



,



> is a

saddle point of L(x,) only when

• x



is a minimum of L(x,



) for each x

 X

• g



)  0 for each i  I





i



) = 0 for each i  I

Theorem. If point <x



,



> is a saddle point of

Lagrangian function L(x,), then x



is an optimum

solution of the NLP

Lagrange Duality (1)

Let L(x,) be a Lagrangian function

L(x,) = f(x) + g(x)

where X  R



= { :   R

,   0}, f : X  R

and g: X  R

Definition. L

(x) is a primal function defined as

follows

(x) = f(x) if g

(x)  0 for each i  I

(x) = + otherwise

Definition. L

() is a dual function defined as

() = inf

xX

L(x,)

Lagrange Duality (2)

Theorem (weak duality). The function value of

the dual at any feasible solution is always lower

than or equal to the function value of the primal

at any feasible solution

(x)  L

()

Definition. Let x



and 



be optimal solutions of the

primal and dual problems, respectively. The

difference (L



) – L

(



)) is called duality gap

Theorem (strong duality). If the objective and

inequality constraints are convex, equality

constraints are linear, then the duality gap is 0

Computational Intelligence

(2)

• Computational intelligence methods are mostly

stochastic, since they use some random

elements

• Computational intelligence algorithms yield

suboptimal results without the optimality

guarantee

• Computational intelligence algorithms solve a

discrete problem

min F(x) dla x  S

• where S is a finite space of feasible solutions

Concluding Remarks

• Many optimization problems are linear with continous

variables and can be solved by Simplex method

• Problems with some integer variables are mostly NP-hard
• To find optimal solution of such problems branch and

bound or branch and cut methods must be used

• Some problems with specific properties can attacked

using classical optimization approaches like duality,

Kuhn-Tucker conditons, Lagrangian relaxation

• Computational intelligence provide effective

algorithms that can be adapted for most of optimization

problems

Document Outline