5.2 Constrained nonlinear optimization
From a mathematical point of view, we have to consider the following minimization task: Minimize f (x )
;
x = x1 L x m
Subject to n equality constraints: h1(x ) = 0
M hn (x ) = 0
Subject to p inequality constraints: g1(x ) ≤ 0
M gp (x ) ≤ 0
5.2.1
5.2 Constrained nonlinear optimization
5.2.2
Introduction of slack variables x m+1...x m+p to transform the inequality constraints into equality constraints g1( x ) + x m2 +1 = 0
M gp ( x ) + x m2 +p = 0
hn+1( x, x m+1 ) = 0
Written in a more general way
M hn+p ( x, x m+p ) = 0
Introduction of the Lagrange function L L( x, ) = f( x) + 1h1( x) + ... +
h ( x) +
n n
h ( x, x m+1) + ... +
n +1 n +1
Necessary conditions for an extremum are: ∂L =0 ∂x
∂L =0 ∂x1
;
x = x1...x m+p ⇒
M ∂L =0 ∂x m+p
h ( x, x m+p )
n +p n +p
5.2 Constrained nonlinear optimization ∂L =0 ∂ 1
and ∂L =0 ∂
5.2.3
;
= 1...
n +p
M ∂L
⇒
∂
The condition
=0
n+p
∂L = 0 applied on the Lagrange function results in ∂x
∂L ∂x1 ∂L = M = ∇L = 0 ∂x ∂L ∂x m+p
(I)
5.2 Constrained nonlinear optimization
The condition
5.2.4
∂L = 0 is identical with the (n + p) constraints ∂
h1( x) = 0
(II)
M hn+p( x) = 0
The necessary conditions leading to equations (I) and (II) can be written in a more compact form as x1 M x m +p z = F( z) = 0 with (III) 1
M n +p
5.2 Constrained nonlinear optimization
5.2.5
Solving equation (III) by an iterative Newton procedure F( z°) +
∂F z + ... = 0 ∂ z z°
Neglecting higher order terms, we obtain ∂F z = −F( z°) ∂ z z°
Iterative procedure for x and the Lagrange multipliers (i)
∇ 2L
G
(i)
(i)
∆x
∇L
(IV)
= GT
0
∆λ
:
h
5.2 Constrained nonlinear optimization
(i)
(i+1)
= λ
(i)
∆x
x
x
5.2.6
+
(V)
λ
∆λ
∇ 2L is the Hessian matrix of the Lagrange function L , G is the functional matrix connected with the constraints h( x ) :
2
∇ L=
∂∇L ∂∇L K ∂x1 ∂x m+p
;
G=
∂h1 ∂hn+p L ∂x1 ∂x1 M M ∂ h ∂h1 L n+ p ∂x m+p ∂x m+p