Calculation of Lagrange multipliers and their use for ...

Calculation of Lagrange multipliers and their use for

sensitivity of optimal solutions

Jaco F. Schutte

EGM 6365 - Structural OptimizationFall 2005

Constrained optimization

x1

x2

Infeasible regions

Feasible region

OptimumDecreasing f(x)

h(x)g(x)

Constraint normalization

0

1 0

a

a

a

g

g

σ σσ σ

σσ

≤= − ≤

= − ≥

Poor optimizer performance often encountered when constraints are not normalized

When normalized g = 0.1 → 10% margin in responce

Equality constraints

� Convert equality constraint to two equality constraints

� Increases the number of constraints

( ) ( )( )

00

0i

ii

h xh x

h x ≤= ⇔ ≥

Reduction of inequality constraints

� Using KS-function (Kreisselmeier-Steinhauser)

� KS bounded by

( )( )

( )

( ) ( )

1

1

00 1 ln

0

jg xj

j

j

g xg x

KS g x e

g x

ρ

ρ−

≥≥ ⇔ = −

≥

∑!

( ) ( )min min

lnj

mg KS g x g

ρ ≤ ≤ −

Using KS to approximate hi(x)

For

the solution lies at hi(x) = - hi(x) = 0

Gradient of KS function of ± hi pair vanishes at solution hi = 0Value of KS function approaches 0 for ρ → ∞

( ) ( )( )

00

0i

ii

h xh x

h x ≤= ⇔ − ≤

( ) ln(2)0 ,KS h hρ

≥ − ≥ −

� Lagrangian function

where λ j are unknown Lagrange multipliers� Stationary point conditions for inequality

constraints:

Kuhn-Tucker conditions

Stationary points

Kuhn-Tucker conditions (contd.)

� Conditions only apply at a regular point (constraint gradients linearly independent)

� Equations

yield n + ne total equations� n stationary point coordinates� ne Lagrange multipliers

� Inequality constraints require transformation to equality constraints:

� This yields the following Lagrangian:

Kuhn-Tucker conditions (contd.)

Kuhn-Tucker conditions (contd.)

� Conditions for stationary points are then:

� If inequality constraint is inactive (t ≠ 0) then Lagrange multiplier = 0

� For inequality constraints a regular point is when gradients of active constraint are linearly independent

Kuhn-Tucker conditions (contd.)

� For an inequality constrained problem x is a minimum of a set of no-negative λ i can be found such that:

1)

2) The corresponding λ i is zero if constraint gj is not active

Kuhn-Tucker conditions (contd.)

Kuhn-Tucker graphical representation

Sufficient conditions� Kuhn-Tucker conditions are satisfied when no

-∇ f can be obtained without violating constraints� Possible to move perpendicular to constraints

and improve objective function (necessary conditions are met, but not sufficient)

� Kuhn tucker conditions for optimality are sufficient when� no. design vars. = no. of active constraints.� or, Hessian of Lagrangian function is positive definite

for subspace tangent to active constraints

Sufficient conditions (contd.)

� For equality constraints

where

� For inequality constraints

A function is convex if

or

Convex problems

� Convex optimization problem has� convex objective function� convex feasible domain if

� All inequality constraints are concave (or �gj = convex)

� All equality constraints are linear

� only one optimum� Kuhn-Tucker conditions necessary and will also

always be sufficient for global minimum

Convex problems (contd.)

Calculating Lagrange multipliers

� Lagrange equation in matrix notation:

where N is

assuming number of active constraints are r define a residual vector u (to be minimized)

Calculating Lagrange multipliers (contd.)

Obtain a least squares solution of u

By differentiating w.r.t. each λ

Or

And substituting into

We obtain where

Calculating Lagrange multipliers (contd.)

� P is a projection matrix which projects a vector into a subspace which is tangent to the constraints

� For the Kuhn Tucker conditions to be satisfied, ∇ f has to be orthogonal to this subspace

� The method of using is often ill-conditioned matrices and inefficient

Alternate method for calculating Lagrange multipliers

QR factorization of N gives a more efficient way ofcalculating λ

Because Q is orthogonal

║u║2 is then minimized by choosing

Alternate method for calculating Lagrange multipliers

Sensitivity of optimum solution to problem parameters

Assuming problem fitness and constraintsdepend on parameter p

The solution is

and the corresponding fitness value

Sensitivity of optimum solution to problem parameters (contd.)

We would like to obtain derivatives of f* w.r.t. p

Equations that govern the optimum solution are

After manipulating governing equations we obtain