Stephen J. Wright
The past ten years have been a time of remarkable developments in software tools for solving optimization problems. There have been algorithmic advances in such areas as linear programming and integer programming which have now borne fruit in the form of more powerful codes. The advent of modeling languages has made the process of formulating the problem and invoking the software much easier, and the explosion in computational power of hardware has made it possible to solve large, difficult problems in a short amount of time on desktop machines. A user community that is growing rapidly in size and sophistication is driving these developments. In this article, we discuss the algorithmic state of the art and its relevance to production codes. We describe some representative software packages and modeling languages and give pointers to web sites that contain more complete information. We also mention computational servers for online solution of optimization problems.
Optimization, Linear programming, Nonlinear programming, Integer programming, Software.
Optimization problems arise naturally in many engineering applications. Control problems can be formulated as optimization problems in which the variables are inputs and states, and the constraints include the model equations for the plant. At successively higher levels, optimization can be used to determine setpoints for optimal operations, to design processes and plants, and to plan for future capacity.
Optimization problems contain the following key ingredients:

Variables that can take on a range of values. Variables that are real numbers, integers, or binary (that is, allowable values 0 and 1) are the most common types, but matrix variables are also possible.

Constraints that define allowable values or scopes for the variables, or that specify relationships between the variables;

An objective function that measures the desirability of a given set of variables.
The optimization problem is to choose from among all variables that satisfy the constraints the set of values that minimizes the objective function.
The term “mathematical programming”, which was coined around 1945, is synonymous with optimization. Correspondingly, linear optimization (in which the constraints and objective are linear functions of the variables) is usually known as “linear programming,” while optimization problems that involve constraints and have nonlinearity present in the objective or in at least some constraints, are known as “nonlinear programming” problems. In convex programming, the objective is a convex function and the feasible set (the set of points that satisfy the constraints) is a convex set. In quadratic programming, the objective is a quadratic function while the constraints are linear. Integer programming problems are those in which some or all of the variables are required to take on integer values.
Optimization technology is traditionally made available to users by means of codes or packages for specific classes of problems. Data is communicated to the software via simple data structures and subroutine argument lists, userwritten subroutines (for evaluating nonlinear objective or constraint functions), text files in the standard MPS format, or text files that describe the problem in certain vendorspecific formats. More recently, modeling languages have become an appealing way to interface to packages, as they allow the user to define the model and data in a way that makes intuitive sense in terms of the application problem. Optimization tools also form part of integrated modeling systems such as GAMS and LINDO, and even underlie spreadsheets such as Microsoft’s Excel. Other “under the hood” optimization tools are present in certain logistics packages, for example, packages for supply chain management or facility location.
The majority of this paper is devoted to a discussion of software packages and libraries for linear and nonlinear programming, both freely available and proprietary. We emphasize in particular packages that have become available during the past 10 years, that address new problem areas or that make use of new algorithms. We also discuss developments in related areas such as modeling languages and automatic differentiation. Background information on algorithms and theory for linear and nonlinear programming can be found in a number of texts, including those of Luenberger (1984), Chvatal (1983), Bertsekas (1995), Nash and Sofer (1996), and the forthcoming book of Nocedal and Wright (1999).
Online Resources and Computational Servers
As with so many other topics, a great deal of information about optimization software is available on the worldwide web. Here we point to a few noncommercial sites that give information about optimization algorithms and software, modeling issues, and operations research. Many other interesting sites can be found by following links from the sites mentioned below.
The NEOS Guide at www.mcs.anl.gov/otc/Guide contains

A guide to optimization software containing around 130 entries. The guide is organized by the name of the code, and classified according to the type of problem solved by the code.

An “optimization tree” containing a taxonomy of optimization problem types and outlines of the basic algorithms.

Case studies that demonstrate the use of algorithms in solving realworld optimization problems. These include optimization of an investment portfolio, choice of a lowestcost diet that meets a set of nutritional requirements, and optimization of a strategy for stockpiling and retailing natural gas, under conditions of uncertainty about future demand and price.
The NEOS Guide also houses the FAQs for Linear and Nonlinear Programming, which can be found at www.mcs.anl.gov/otc/Guide/faq/. These pages, updated monthly, contain basic information on modeling and algorithmic issues, information for most of the available codes in the two areas, and pointers to texts for readers who need background information.
Michael Trick maintains a comprehensive web site on operations research topics at http://mat.gsia.cmu.edu. It contains pointers to most online resources in operations research, together with an extensive directory of researchers and research groups and of companies that are involved in optimization and logistics software and consulting.
Hans Mittelmann and Peter Spellucci maintain a decision tree to help in the selection of appropriate optimization software tools at http://plato.la.asu.edu/guide.html. Benchmarks for a variety of codes, with an emphasis on linear programming solvers that are freely available to researchers, can be found at http://plato.la.asu.edu/bench.html. The page http://solon.cma.univie.ac.at/~neum/glopt.html, maintained by Arnold Neumaier, emphasizes global optimization algorithms and software.
The NEOS Server at www.mcs.anl.gov/neos/Server is a computational server for the remote solution of optimization problems over the Internet. By using an email interface, a Web page, or an xwindows “submission tool” that connects directly to the Server via Unix sockets, users select a code and submit the model information and data that define their problem. The job of solving the problem is allocated to one of the available workstations in the Server’s pool on which that particular package is installed, then the problem is solved and the results returned to the user.
The Server now has a wide variety of solvers in its roster, including a number of proprietary codes. For linear programming, the BPMPD, HOPDM, PCx, and XPRESSMP/BARRIER interiorpoint codes as well as the XPRESSMP/SIMPLEX code are available. For nonlinear programming, the roster includes LANCELOT, LOQO, MINOS, NITRO, SNOPT, and DONLP2. Input in the AMPL modeling language is accepted for many of the codes.
The obvious target audience for the NEOS Server includes users who want to try out a new code, to benchmark or compare different codes on data of relevance to their own applications, or to solve small problems on an occasional basis. At a higher level, however, the Server is an experiment in using the Internet as a computational, problemsolving tool rather than simply an informational device. Instead of purchasing and installing a piece of software for installation on their local hardware, users gain access to the latest algorithmic technology (centrally maintained and updated), the hardware resources needed to execute it and, where necessary, the consulting services of the authors and maintainers of each software package. Such a means of delivering problemsolving technology to its customers is an appealing option in areas that demand access to huge amounts of computing cycles (including, perhaps, integer programming), areas in which extensive handson consulting services are needed, areas in which access to large, centralized, constantly changing data bases, and areas in which the solver technology is evolving rapidly.
L
inear Programming
In linear programming problems, we minimize a linear function of real variables over a region defined by linear constraints. The problem can be expressed in standard form as
w
here x is a vector of n real numbers, while is a set of linear equality constraints and indicates that all components of x are required to be nonnegative. The dual of this problem is
where is a vector of Lagrange multipliers and is a vector of dual slack variables. These two problems are intimately related, and algorithms typically solve both of them simultaneously. When the vectors and satisfy the following optimality conditions:
then solves the primal problem and solves the dual problem.
Simple transformations can be applied to any problem with a linear objective and linear constraints (equality and inequality) to obtain this standard form. Production quality linear programming solvers carry out the necessary transformations automatically, so the user is free to specify upper bounds on some of the variables, use linear inequality constraints, and in general make use of whatever formulation is most natural for their particular application.
The popularity of linear programming as an optimization paradigm stems from its direct applicability to many interesting problems, the availability of good, generalpurpose algorithms, and the fact that in many realworld situations, the inexactness in the model or data means that the use of a more sophisticated nonlinear model is not warranted. In addition, linear programs do not have multiple local minima, as may be the case with nonconvex optimization problems. That is, any local solution of a linear programone whose function value is no larger than any feasible point in its immediate vicinityalso achieves the global minimum of the objective over the whole feasible region. It remains true that more (human and computational) effort is invested in solving linear programs than in any other class of optimization problems.
Prior to 1987, all of the commercial codes for solving general linear programs made use of the simplex algorithm. This algorithm, invented in the late 1940s, had fascinated optimization researchers for many years because its performance on practical problems is usually far better than the theoretical worst case. A new class of algorithms known as interiorpoint methods was the subject of intense theoretical and practical investigation during the period 1984—1995, with practical codes first appearing around 1989. These methods appeared to be faster than simplex on large problems, but the advent of a serious rival spurred significant improvements in simplex codes. Today, the relative merits of the two approaches on any given problem depend strongly on the particular geometric and algebraic properties of the problem. In general, however, good interiorpoint codes continue to perform as well or better than good simplex codes on larger problems when no prior information about the solution is available. When such “warm start” information is available, however, as is often the case in solving continuous relaxations of integer linear programs in branchandbound algorithms, simplex methods are able to make much better use of it than interiorpoint methods. Further, a number of good interiorpoint codes are freely available for research purposes, while the few freely available simplex codes are not quite competitive with the best commercial codes.
The simplex algorithm generates a sequence of feasible iterates for the primal problem, where each iterate typically has the same number of nonzero (strictly positive) components as there are rows in . We use this iterate to generate dual variables and such that two other optimality conditions are satisfied, namely,
If the remaining condition is also satisfied, then the solution has been found and the algorithm terminates. Otherwise, we choose one of the negative components of and allow the corresponding component of to increase from zero. To maintain feasibility of the equality constraint the components that were strictly positive in will change. One of them will become zero when we increase the new component to a sufficiently large value. When this happens, we stop and denote the new iterate by .
Each iteration of the simplex method is relatively inexpensive. It maintains a factorization of the submatrix of that corresponds to the strictly positive components of a square matrix known as the basisand updates this factorization at each step to account for the fact that one column of has changed. Typically, simplex methods converge in a number of iterates that is about two to three times the number of columns in .
Interiorpoint methods proceed quite differently, applying a Newtonlike algorithm to the three equalities in the optimality conditions and taking steps that maintain strict positivity of all the and components. It is the latter feature that gives rise to the term “interiorpoint” the iterates are strictly interior with respect to the inequality constraints. Each interiorpoint iteration is typically much more expensive than a simplex iteration, since it requires refactorization of a large matrix of the form , where and are diagonal matrices whose diagonal elements are the components of the current iterates and , respectively. The solutions to the primal and dual problems are generated simultaneously. Typically, interiorpoint iterates converge in between 10 and 100 iterations.
Codes can differ in a number of important respects, apart from the different underlying algorithm. All practical codes include presolvers, which attempt to reduce the dimension of the problem by determining the values of some of the primal and dual variables without applying the algorithm. As a simplex example, suppose that the linear program contains the constraints
then the only possible values for the three variables are
These variables can be fixed and deleted from the problem, along with the three corresponding columns of and the three components of . Presolve techniques have become quite sophisticated over the years, though little has been written about them because of their commercial value. An exception is the paper of Andersen and Andersen (1995).
For information on specific codes, refer to the online resources mentioned earlier; in particular, the NEOS Software Guide, the Linear Programming FAQ, and the benchmarks maintained by Hans Mittelmann.
Modern, widely used commercial simplex codes include CPLEX and the XPRESSMP. Both these codes accept input in the industrystandard MPS format, and also in their own proprietary formats. Both have interfaces to various modeling languages, and also a “callable library” interface that allows users to set up, modify, and solve problems by means of function calls from C or FORTRAN code. Both packages are undergoing continual development. Freely available simplex codes are usually of lower quality, with the exception of SOPLEX. This is a C++ code, written as a thesis project by Roland Wunderling, that can be found at www.zib.de/Optimization/Software/Soplex/. The code MINOS is available to nonprofit and academic researchers for a nominal fee.
Commercial interiorpoint solvers are available as options in the CPLEX and XPRESSMP packages. However, a number of highly competitive codes are available free for research and noncommercial use, and can for the most part be obtained through the Web. Among these are BPMPD, PCx, COPLLP, LOQO, HOPDM, and LIPSOL. See Mittelmann’s benchmark page for comparisons of these codes and links to their web sites. These codes mostly charge a license fee for commercial use, but it is typically lower than for fully commercial packages. All can read MPS files, and most are interfaced to modeling languages. LIPSOL is programmed in Matlab (with the exception of the linear equations solver), while the other codes are written in C and/or FORTRAN.
A fine reference on linear programming, with an emphasis on the simplex method, is the book of Chvatal (1983). An online Java applet that demonstrates the operation of the simplex method on small userdefined problems can be found at www.mcs.anl.gov/otc/Guide/CaseStudies/simplex/. Wright (1997) gives a description of practical interiorpoint methods.
Modeling Languages
From the user’s point of view, the efficiency of the algorithm or the quality of the programming may not be the critical factors in determining the usefulness of the code. Rather, the ease with which it can be interfaced to his particular applications may be more important; weeks of personhours may be more costly to the enterprise than a few hours of time on a computer. The most suitable interface depends strongly on the particular application and on the context in which it is solved. For users that are well acquainted with a spreadsheet interface, for instance, or with MATLAB, a code that can accept input from these sources may be invaluable. For users with large legacy modeling codes that set up and solve optimization problems by means of subroutine calls, substitution of a more efficient package that uses more or less the same subroutine interface may be the best option. In some disciplines, (JP’s biology/chemistry pointer) applicationspecific modeling languages allow problems to be posed in a thoroughly intuitive way. In other cases, applicationspecific graphical user interfaces may be more appropriate.
For general optimization problems, a number of highlevel modeling languages have become available that allow problems to be specified in intuitive terms, using data structures, naming schemes, and algebraic relational expressions that are dictated by the application and model rather than by the input requirements of the optimization code. Typically, a user starting from scratch will find the process of model building more straightforward and bug free with such a modeling language than, say, a process of writing FORTRAN code to pack the data into onedimensional arrays, turning the algebraic relations between the variables into FORTRAN expressions involving elements of these arrays, and writing more code to interpret the output from the optimization routine in terms of the original application.
The following simple example in AMPL demonstrates the usefulness of a modeling language (see Fourer, Gay, and Kernighan (1993), page 11). The application is to a steel production model, in which the aim is to maximize profit obtained from manufacturing a number of steel products by choosing the amount of each product to manufacture, subject to restrictions on the maximum demands for each product and the time available in each work week to manufacture them. The following file is an AMPL “model file” that specifies the variables, the parameters that quantify aspects of the model, and the constraints and objective.
set PROD;
param rate {PROD} >0;
param avail >= 0;
param profit {PROD};
param market{PROD};
var Make {p in PROD} >= 0, <= market[p];
maximize total_profit: sum {p in PROD} profit[p] *Make[p];
subject to Time: sum {p in PROD} (1/rate[p]) * Make[p] <= avail;
PROD is the collection of possible products that can be manufactured, while rate, profit and market are the rate at which each product can be manufactured, the profit on each product, and the maximum demand for each product, respectively. avail represents the total time available for manufacturing. Make is the variable in the problem, representing the amount of each product to be manufactured. In its definition, each element of Make is constrained to lie between zero and the maximum demand for the product in question. The last two lines of the model file specify the objective and constraint in a selfevident fashion.
The actual values of the parameters can be assigned by means of additional statements in this file, or in a separate “data file.” For instance, the following data file specifies parameters for two products, bands and coils:
set PROD := bands coils;
param: rate profit market :=
bands 200 25 6000
coils 140 30 4000;
param avail := 40;
These statements specify that the market[bands] is 6000, profit[bands] is 25, and so on. An interactive AMPL session would proceed by invoking commands to read these two files and then invoking an option solver command to choose the linear programming solver to be used (for example, CPLEX or PCx) together with settings for parameters such as stopping tolerances, etc, that the user may wish to change from their defaults. A solve command would then solve the problem (and report messages passed through from the underlying optimization code). Results can be inspected by invoking the display command. For the above example, the command display Make invoked after the problem has been solved would produce the following output:
Make [*] :=
bands 6000
coils 0
;
Note from this example the intuitive nature of the algebraic relations, and the fact that we could index the parameter arrays by the indices bands and coils, rather than the numerical indices 1 and 2 that would be required if we were programming in FORTRAN. Note too that additional products can be added to the mix without changing the model file at all.
Of course, the features of AMPL are much more extensive than the simple example above allows us to demonstrate. The web site www.ampl.com contains a great deal of information about the language and the optimization software to which it is linked, and allows users to solve their own simple models online.
Numerous other modeling languages and systems can be found on the online resources described above, particularly the NEOS Software Guide and the linear and nonlinear programming FAQ’s. We mention in particular AIMMS (Bisschop and Entriken (1993)) which has a built in graphical interface; GAMS (www.gams.com), a well established system available with support for linear, nonlinear, and mixedinteger programming and newly added procedural features; and MPL, a Windowsbased system whose web site www.maximalusa.com contains a comprehensive tutorial and a free student version of the language.