This section shows how to choose whether to convert a nonlinear function to an optimization expression or to create the expression out of supported operations on optimization variables. The section also shows how to convert a function, if necessary, by using
Generally, create your objective or nonlinear constraint functions by using supported operations on optimization variables and expressions. Doing so has these advantages:
solve includes gradients calculated by automatic differentiation. See Effect of Automatic Differentiation in Problem-Based Optimization.
solve has a wider choice of available solvers. When using
solve uses only
In general, supported operations include all elementary mathematical operations: addition, subtraction, multiplication, division, powers, and elementary functions such as exponential and trigonometric functions and their inverses. Nonsmooth operations such as
case are not supported. For the complete description, see Supported Operations for Optimization Variables and Expressions.
For example, suppose that your objective function is
where is a parameter that you supply, and the problem is to minimize over and . This objective function is a sum of squares, and takes the minimal value of 0 at the point , .
The objective function is a polynomial, so you can write it in terms of elementary operations on optimization variables.
r = 2; x = optimvar('x'); y = optimvar('y'); f = 100*(y - x^2)^2 + (r - x)^2; prob = optimproblem("Objective",f); x0.x = -1; x0.y = 2; [sol,fval] = solve(prob,x0)
Solving problem using lsqnonlin. Local minimum found. Optimization completed because the size of the gradient is less than the value of the optimality tolerance.
sol = struct with fields: x: 2.0000 y: 4.0000
fval = 8.0661e-29
To solve the same problem by converting the objective function using
fcn2optimexpr (not recommended), first write the objective as an anonymous function.
fun = @(x,y)100*(y - x^2)^2 + (r - x)^2; prob.Objective = fcn2optimexpr(fun,x,y); [sol2,fval2] = solve(prob,x0)
Solving problem using fminunc. Local minimum found. Optimization completed because the size of the gradient is less than the value of the optimality tolerance.
sol2 = struct with fields: x: 2.0000 y: 3.9998
fval2 = 1.7143e-09
fminunc this time instead of the more efficient
lsqnonlin, and the reported solution for
y is slightly different than the correct solution 4. Furthermore, the reported
fval is about 1e-9 instead of 1e-20 (the actual solution value is exactly 0). These slight inaccuracies are due to
solve not using the more efficient solver.
The remainder of this example shows how to convert a function to an optimization expression using
To use a function file in the problem-based approach, you must convert the file to an expression using
For example, the
expfn3.m file contains the following code:
function [f,g,mineval] = expfn3(u,v) mineval = min(eig(u)); f = v'*u*v; f = -exp(-f); t = u*v; g = t'*t + sum(t) - 3;
This function is not entirely composed of supported operations because of
min(eig(u)). Therefore, to use
expfn3(u,v) as an optimization expression, you must first convert it using
expfn3 as an optimization expression, first create optimization variables of the appropriate sizes.
u = optimvar('u',3,3,'LowerBound',-1,'UpperBound',1); % 3-by-3 variable v = optimvar('v',3,'LowerBound',-2,'UpperBound',2); % 3-by-1 variable
Convert the function file to an optimization expressions using
[f,g,mineval] = fcn2optimexpr(@expfn3,u,v);
Because all returned expressions are scalar, you can save computing time by specifying the expression sizes using the
'OutputSize' name-value pair argument. Also, because
expfn3 computes all of the outputs, you can save more computing time by using the
ReuseEvaluation name-value pair.
[f,g,mineval] = fcn2optimexpr(@expfn3,u,v,'OutputSize',[1,1],'ReuseEvaluation',true)
f = Nonlinear OptimizationExpression [argout,~,~] = expfn3(u, v)
g = Nonlinear OptimizationExpression [~,argout,~] = expfn3(u, v)
mineval = Nonlinear OptimizationExpression [~,~,argout] = expfn3(u, v)
To use a general nonlinear function handle in the problem-based approach, convert the handle to an optimization expression using
fcn2optimexpr. For example, write a function handle equivalent to
mineval and convert it.
fun = @(u)min(eig(u)); funexpr = fcn2optimexpr(fun,u,'OutputSize',[1,1])
funexpr = Nonlinear OptimizationExpression anonymousFunction2(u) where: anonymousFunction2 = @(u)min(eig(u));
To use the objective expression as an objective function, create an optimization problem.
prob = optimproblem; prob.Objective = f;
Define the constraint
g <= 0 in the optimization problem.
prob.Constraints.nlcons1 = g <= 0;
Also define the constraints that
u is symmetric and that .
prob.Constraints.sym = u == u.'; prob.Constraints.mineval = mineval >= -1/2;
View the problem.
OptimizationProblem : Solve for: u, v minimize : [argout,~,~] = expfn3(u, v) subject to nlcons1: arg_LHS <= 0 where: [~,arg_LHS,~] = expfn3(u, v); subject to sym: u(2, 1) - u(1, 2) == 0 u(3, 1) - u(1, 3) == 0 -u(2, 1) + u(1, 2) == 0 u(3, 2) - u(2, 3) == 0 -u(3, 1) + u(1, 3) == 0 -u(3, 2) + u(2, 3) == 0 subject to mineval: arg_LHS >= (-0.5) where: [~,~,arg_LHS] = expfn3(u, v); variable bounds: -1 <= u(1, 1) <= 1 -1 <= u(2, 1) <= 1 -1 <= u(3, 1) <= 1 -1 <= u(1, 2) <= 1 -1 <= u(2, 2) <= 1 -1 <= u(3, 2) <= 1 -1 <= u(1, 3) <= 1 -1 <= u(2, 3) <= 1 -1 <= u(3, 3) <= 1 -2 <= v(1) <= 2 -2 <= v(2) <= 2 -2 <= v(3) <= 2
To solve the problem, call
solve. Set an initial point
rng default % For reproducibility x0.u = 0.25*randn(3); x0.u = x0.u + x0.u.'; x0.v = 2*randn(3,1); [sol,fval,exitflag,output] = solve(prob,x0)
Solving problem using fmincon. Feasible point with lower objective function value found. Local minimum found that satisfies the constraints. Optimization completed because the objective function is non-decreasing in feasible directions, to within the value of the optimality tolerance, and constraints are satisfied to within the value of the constraint tolerance.
sol = struct with fields: u: [3x3 double] v: [3x1 double]
fval = -403.4288
exitflag = OptimalSolution
output = struct with fields: iterations: 87 funcCount: 1448 constrviolation: 6.3860e-12 stepsize: 7.4093e-05 algorithm: 'interior-point' firstorderopt: 0.0012 cgiterations: 172 message: '...' bestfeasible: [1x1 struct] objectivederivative: "finite-differences" constraintderivative: "finite-differences" solver: 'fmincon'
View the solution.
0.8419 0.5748 -0.7670 0.5748 0.3745 0.2997 -0.7670 0.2997 0.5667
2.0000 -2.0000 2.0000
The solution matrix
u is symmetric. All values of
v are at the bounds.
Copyright 2018–2020 The MathWorks, Inc.