Write Constraints

Consult Optimization Toolbox Documentation

Many Global Optimization Toolbox functions accept bounds, linear constraints, or nonlinear constraints. To see how to include these constraints in your problem, see Write Constraints. Try consulting these pertinent links to sections:

Note

The surrogateopt solver uses a different syntax for nonlinear constraints than other solvers, and requires finite bounds on all components. For details, see the function reference page and Convert Nonlinear Constraints Between surrogateopt Form and Other Solver Forms.

Set Bounds

It is more important to set bounds for global solvers than for local solvers. Global solvers use bounds in a variety of ways:

GlobalSearch requires bounds for its scatter-search point generation. If you do not provide bounds, GlobalSearch bounds each component below by -9999 and above by 10001. However, these bounds can easily be inappropriate.
If you do not provide bounds and do not provide custom start points, MultiStart bounds each component below by -1000 and above by 1000. However, these bounds can easily be inappropriate.
ga uses bounds and linear constraints for its initial population generation. For unbounded problems, ga uses a default of 0 as the lower bound and 1 as the upper bound for each dimension for initial point generation. For bounded problems, and problems with linear constraints, ga uses the bounds and constraints to make the initial population.
simulannealbnd and patternsearch do not require bounds, although they can use bounds.

Ensure ga Options Maintain Feasibility

The ga solver generally maintains strict feasibility with respect to bounds and linear constraints. This means that, at every iteration, all members of a population satisfy the bounds and linear constraints.

However, you can set options that cause this feasibility to fail. For example if you set MutationFcn to @mutationgaussian or @mutationuniform, the mutation function does not respect constraints, and your population can become infeasible. Similarly, some crossover functions can cause infeasible populations, although the default gacreationlinearfeasible does respect bounds and linear constraints. Also, ga can have infeasible points when using custom mutation or crossover functions.

To ensure feasibility, use the default crossover and mutation functions for ga. Be especially careful that any custom functions maintain feasibility with respect to bounds and linear constraints.

Note

When a problem has integer constraints, ga ensures that all operators (mutation, crossover, and creation) return feasible populations with respect to bounds, linear constraints, and integer constraints at each iteration. This feasibility holds to within a small tolerance.

Gradients and Hessians

If you use GlobalSearch or MultiStart with fmincon, your nonlinear constraint functions can return derivatives (gradient or Hessian). For details, see Gradients and Hessians.

Vectorized Constraints

The ga and patternsearch solvers optionally compute the nonlinear constraint functions of a collection of vectors in one function call. This method can take less time than computing the objective functions of the vectors serially. This method is called a vectorized function call.

For the solver to compute in a vectorized manner, you must vectorize both your objective (fitness) function and nonlinear constraint function. For details, see Vectorize the Objective and Constraint Functions.

As an example, suppose your nonlinear constraints for a three-dimensional problem are

$\begin{matrix} \frac{x_{1}^{2}}{4} + \frac{x_{2}^{2}}{9} + \frac{x_{3}^{2}}{25} \leq 6 \\ x_{3} \geq \cosh (x_{1} + x_{2}) \\ x_{1} x_{2} x_{3} = 2. \end{matrix}$

The following code gives these nonlinear constraints in a vectorized fashion, assuming that the rows of your input matrix x are your population or input vectors:

function [c ceq] = nlinconst(x)

c(:,1) = x(:,1).^2/4 + x(:,2).^2/9 + x(:,3).^2/25 - 6;
c(:,2) = cosh(x(:,1) + x(:,2)) - x(:,3);
ceq = x(:,1).*x(:,2).*x(:,3) - 2;

For example, minimize the vectorized quadratic function

function y = vfun(x)
y = -x(:,1).^2 - x(:,2).^2 - x(:,3).^2;

over the region with constraints nlinconst using patternsearch:

options = optimoptions('patternsearch','UseCompletePoll',true,'UseVectorized',true);
[x fval] = patternsearch(@vfun,[1,1,2],[],[],[],[],[],[],...
    @nlinconst,options)

Optimization terminated: mesh size less than options.MeshTolerance
 and constraint violation is less than options.ConstraintTolerance.

x =
    0.2191    0.7500   12.1712

fval =
 -148.7480

Using ga:

options = optimoptions('ga','UseVectorized',true);
[x fval] = ga(@vfun,3,[],[],[],[],[],[],@nlinconst,options)

Optimization terminated: maximum number of generations exceeded.

x =
   -1.4098   -0.1216   11.6664

fval =
 -138.1066

For this problem patternsearch computes the solution far more quickly and accurately.