Nonlinear Constrained Optimization
Nonlinear Constrained Optimization
J ? = min f (x)
s.t. x 2 Rn
Gradient
available?
yes
Michal Kvasnica
Linear
equations?
yes
f = 0
Gradient-Free Methods
Gradientfree mth.
no
Gradientbased mth.
Constrained Optimization
J ? = min f (x)
Random search:
s.t. g(x) 0,
- Luus-Jaakola
h(x) = 0
- simulated annealing
no
Assumptions:
- functions f, g, h are non-convex functions
KKT method
- only necessary conditions
- exponential complexity
J ? = min f (x)
J ? = min f (x)
s.t. g(x) 0,
s.t. g(x) 0,
h(x) = 0
h(x) = 0
Idea:
Idea:
xk ) + 1/2 (x
xk )
s.t. g(x) 0,
h(x) = 0
xk ) + 1/2 (x
xk )
J ? = min f (x)
s.t. g(x) 0,
xk ) T
J ? = min f (x)
xk )T r2 f (xk )(x
xk )T r2 f (xk )(x
h(x) = 0
xk ) T
xk ) 0,
xk ) = 0
xk ) + 1/2 (x
xk )T r2 f (xk )(x
xk ) T
xk ) 0,
xk ) = 0
Iterations:
- start with a feasible x0
- solve the QP approximation, obtain x*
- set xk+1 = x* and iterate until convergence
xk ) + 1/2 (x
xk ) 0,
xk ) = 0
xk )T (I + r2 f (xk ))(x
xk ) T
SQP: Example
SQP: Example
s.t.
3x2
2x + 5
3x2
s.t.
SQP: Example
3x2
2x + 5
3x2
2x + 5
3x2
SQP: Example
s.t.
3x2
3x2
2x + 5
s.t.
3x2
SQP: Example
SQP: Example
s.t.
3x2
2x + 5
3x2
s.t.
SQP: Example
3x2
2x + 5
3x2
2x + 5
3x2
SQP: Example
s.t.
3x2
3x2
2x + 5
s.t.
3x2
SQP: Example
SQP: Example
s.t.
3x2
2x + 5
3x2
s.t.
SQP: Example
3x2
2x + 5
3x2
2x + 5
3x2
SQP: Example
s.t.
3x2
3x2
2x + 5
s.t.
3x2
SQP: Example
min 0.5x4 + 0.8x3
x
s.t.
3x2
3x2
2x + 5