next up previous contents index CD CD Algorithms
Next: Linear Programming Up: Numerical Problems Previous: Determinants and Permanents

Constrained and Unconstrained Optimization



Input description: A function tex2html_wrap_inline27200 .

Problem description: What point tex2html_wrap_inline27202 maximizes (or minimizes) the function f?

Discussion: Most of this book concerns algorithms that optimize one thing or another. This section considers the general problem of optimizing functions where, due to lack of structure or knowledge, we are unable to exploit   the problem-specific algorithms seen elsewhere in this book.

Optimization arises whenever there is an objective function that must be tuned for optimal performance.   Suppose we are building a program to identify good stocks to invest in.   We have available certain financial data to analyze, such as the price-earnings ratio, the interest and inflation rates, and the stock price, all as a function of time t. The key question is how much weight we should give to each of these factors, where these weights correspond to coefficents of a formula:


We seek the numerical values tex2html_wrap_inline27204 , tex2html_wrap_inline27206 , tex2html_wrap_inline27208 , tex2html_wrap_inline27210 whose stock-goodness function does the best job of evaluating stocks.    Similar issues arise in tuning evaluation functions for game playing programs such as chess.

Unconstrained optimization problems also arise in scientific computation.   Physical systems from protein structures to particles naturally seek to minimize their ``energy functions.'' Thus programs that attempt to simulate nature often define energy potential functions for the possible configurations of objects and then take as the ultimate configuration the one that minimizes this potential.   

Global optimization problems tend to be hard, and there are lots of ways to go about them.   Ask the following questions to steer yourself in the right direction:

Efficient algorithms for unconstrained global optimization use derivatives and partial derivatives to find local optima, to point out the direction in which moving from the current point does the most to increase or decrease the function. Such derivatives can sometimes be computed analytically, or they can be estimated numerically by taking the difference between values of nearby points. A variety of steepest descent and conjugate gradient methods to find local optima have been developed, similar in many ways to numerical root-finding algorithms.        

It is a good idea to try out several different methods on any given optimization problem. For this reason, we recommend experimenting with the implementations below before attempting to implement your own method. Clear descriptions of these algorithms are provided in several numerical algorithms books, in particular [PFTV86].

For constrained optimization, finding points that satisfy all the constraints is often the difficult problem. One approach is to use a method for unconstrained optimization, but add a penalty according to how many constraints are violated.     Determining the right penalty function is problem-specific, but it often makes sense to vary the penalties as optimization proceeds. At the end, the penalties should be very high to ensure that all constraints are satisfied.

Simulated annealing is a fairly robust and simple approach to constrained optimization, particularly when we are optimizing over combinatorial   structures (permutations, graphs, subsets) instead of continuous functions. Techniques for simulated annealing are described in Section gif.

Implementations: Several of the Collected Algorithms of the ACM are Fortran codes for unconstrained optimization, most notably Algorithm 566 [MGH81], Algorithm 702 [SF92], and Algorithm 734 [Buc94]. Algorithm 744 [Rab95] does unconstrained optimization in Lisp. They are available from Netlib (see Section gif).   Also check out the selection at GAMS, the NIST Guide to Available Mathematical Software, at    

NEOS (Network-Enabled Optimization System) provides a unique service, the opportunity to solve your problem on computers and software at Argonne National Laboratory, over the WWW.     Linear programming and unconstrained optimization are both supported. This is worth checking out at when you need a solution instead of a program.

General purpose simulated annealing implementations are available and probably are the best place to start experimenting with this technique for constrained optimization.     Particularly popular is Adaptive Simulated Annealing (ASA), written in C and retrievable via anonymous ftp from [] in the /pub/ingber directory.   To get on the ASA mailing list send e-mail to

Genocop, by Zbigniew Michalewicz [Mic92], is a genetic algorithm-based program for constrained and unconstrained optimization, written in C.     I tend to be quite skeptical of genetic algorithms (see Section gif), but many people find them irresistible. Genocop is available from for noncommercial purposes.

Notes: Steepest-descent methods for unconstrained optimization are discussed in most books on numerical methods, including [PFTV86, BT92]. Unconstrained optimization is the topic of several books, including [Bre73, Fle80].

Simulated annealing was devised by Kirkpatrick et. al. [KGV83] as a modern variation of the Metropolis algorithm [MRRT53].       Both use Monte Carlo techniques to compute the minimum energy state of a system. Good expositions on simulated annealing include [AK89].

Genetic algorithms were developed and popularized by Holland [Hol75, Hol92].   Expositions on genetic algorithms include [Gol89, Koz92, Mic92]. Tabu search [Glo89a, Glo89b, Glo90]   is yet another heuristic search procedure with a devoted following.

Related Problems: Linear programming (see page gif), satisfiability (see page gif).    

next up previous contents index CD CD Algorithms
Next: Linear Programming Up: Numerical Problems Previous: Determinants and Permanents

Mon Jun 2 23:33:50 EDT 1997