Chapter 4 Constrained Optimization

advertisement
Chapter 4
Constrained Optimization
 Introduction
The standard form of the constrained optimization problem is as follows:
Minimize f ( x )
Subject to
gu ( x)  0
hv ( x)  0
x  Rn
u  1 , 2 , .m. . ,
v  1 , 2 , . p. ., n
According to the solve algorithm, indirect method and direct method. Direct method: Sequential
Linear Programming; Generalized Reduced gradient method, Methods of feasible direction.
Indirect method: Penalty function method, Interior and exterior penalty function method.
Direct search methods persist for several good reasons. First and foremost, direct search
methods have remained popular because they work well in practice. In fact, many of the direct
search methods are based on surprisingly sound heuristics(启发) that fairly recent analysis
demonstrates guarantee global convergence behavior analogous to the results known for
globalized quasi-Newton techniques. Direct search methods succeed because many of them can be
shown to rely on techniques of classical analysis in ways that are not readily apparent from their
original specifications.
Second, quasi-Newton methods are not applicable to all nonlinear optimization problems.
Direct search methods have succeeded when more elaborate approaches failed. Features unique to
direct search methods often avoid the pitfalls(缺陷) that can plague(苦恼) more sophisticated
approaches.
Third, direct search methods can be the method of first recourse, even among well-informed
users. The reason is simple enough: direct search methods are reasonably straight forward to
implement and can be applied almost immediately to many nonlinear optimization problems. The
requirements from a user are minimal and the algorithms themselves require the setting of few
parameters. It is not unusual for complex optimization problems to require further software
development before quasi-Newton methods can be applied. For such problems, it can make sense
to begin the search for a minimizer using a direct search method with known global convergence
properties, while undertaking the preparations for the quasi-Newton method. When the
preparations for the quasi-Newton method have been completed, the best known result from the
direct search calculation can be used as a “hot start” for one of the quasi-Newton approaches,
which enjoy superior local convergence properties. Such hybrid optimization strategies are as old
as the direct search methods themselves.
Random Search Method:
Generates trial solution for the decision variables.
Classification: Random jump, random walk, and random walk with direction exploitation.
Random Jump: generates huge number of data points assuming uniform distribution of decision
variables and selects the best one.
Random Walk: generates trial solution with sequential improvements using scalar step length and
unit random vector.
Random Walk with Direction Exploitation: Improved version of random walk search, successful
direction of generating trial solution is found out and steps are taken along this direction.
Univariate search
Perform search in one direction at a time using one-dimensional search methods.Go through all
variables in a sequence and then repeat the sequence.

Penalty function methods
Penalty function methods are used to convert constrained problems into unconstrained problems.
In this class of methods we replace the original constrained problem by a sequence of
unconstrained subproblems that minimizes the penalty functions. The penalty function is a
function with penalty property
m
l
j 1
k 1
  x, r1 , r2   f  x   r1  G  g j  x   r2  H hk  x 
Constructed from the objective function f ( x ) and the constraints g , h . The so-called
“penalty” property requires  ( x)  f ( x) for all feasible points x ∈ X , and  ( x ) is much larger
than f ( x ) when the constraint violations are severe.

Interior Penalty Method
If the penalty function takes values approaching +∞ as x approaches the boundary of the
feasible region, it is called the interior point penalty function. The interior point penalty function is
suitable only to inequality-constrained problems. Typically, the two most important interior point
penalty functions are the inverse barrier function:
m
  x, r   f  x   r 
j 1
1
g j  x
and the logarithmic barrier function
m
  x, r   f  x   r  ln  g j  x 
j 1
where r reduces from a high value to 0 gradually.
If given an initial point in the interior of the feasible region, the whole sequence generated by the
interior point penalty function method is interior points. Since these functions set an infinitely
high “barrier” on the boundary, they are also said to be barrier functions.
Selecting r at the start of the optimization procedure:
r1  ( 0.1 ~ 1 ) 
f ( x1 )
where x1 is the initial point in the feasible region.
1

g ( x1 )
Flow-chart summarizes the optimization procedure of the interior penalty method. Here the
reduction follows:
rk 1  c  rk
where c is a coefficient less than 1.
Hence ordinary penalty functions generally require the value of some coefficients to be specified
at the beginning of optimization. However these coefficients usually have no clear physical
meaning. Consequently, it is very difficult to select appropriate values for these coefficients even
by experience.

Exterior Penalty Method
The widely used form of  in the exterior penalty method is:
m
l
  x, r   f  x   r  max 0, g j  x    r   hk  x  
2
j 1
2
k 1
and r is a parameter which is modified at the beginning of each round of optimization. Each
r
optimization round is defined here as a complete optimization of  (x, r ) for a fixed value of k
until the convergence is achieved. The optimum point x
*
at the end of each round serves as the
r
starting point x1 of the next round of optimization with a larger k . Flow-chart shows the general
optimization procedure of the exterior penalty method for problems with inequality and equality
constraints.
The selection of appropriate
rk values is vital for faster convergence and more precision. In
some cases the user might specify the value of
rk at the end of each optimization round. This
technique is very interactive and time consuming and is not generally preferred. Another dominant
technique is to define a function which automatically determines the
rk value at the beginning of
a new round of optimization. Denoting the optimization steps by k, then k  1 at the onset of
optimization and
rk  r1 . The selection of an appropriate r1 plays a key role in the convergence
behavior of the method. They have also suggested that rk be updated according to: rk 1  c rk
where for most structural problems a value of c  5 has been found satisfactory. The third
alternative to determine appropriate rk is to use the so called intelligent and adaptive techniques
such as fuzzy logic and neural networks.
Download