yalmip / yalmip

MATLAB toolbox for optimization modeling

YALMIP

MATLAB toolbox for optimization modeling

Official homepage https://yalmip.github.io

yalmip's Issues

checkset reports strange indicator for binary

Should report 0 since feasible
z = sdpvar(2,1);
assign(z,[0;1]);
checkset(binary(z))

Add property whether solver supports initial guess

Make it granular, such as no, partial solution, complete solution

Warn if user tries to assign initial solution without support

Plot of low-dimensional sets

Fails

sdpvar x1 x2 u
plot([-1 <= x1 <= 1, x1+x2 == 1, 0 <= u <= 1])

cumsum fails

dx = sdpvar(15,1);
cumsum(dx)

Exploit symmetry in matrix norm

The 1-norm case should exploit symmetry when defining auxilliary variables and constraints

Convexity propagation fail

Should not work without binary variables, but it does...

 sdpvar x
 solvesdp([-1 <= x <= 1],x*abs(x))

Bug in robustify

line #124 should be changed from

[F_eq_left,F_eliminate_equality] = filter_eliminatation(F_eq,w,0);

to

[F_eq_left,F_eliminate_equality] = filter_eliminatation(F_eq,w,0, ops);

Check internal structs for consitency

Add a time tag when the internals are cleared and attach this to all objects created. If time tag on object is older than in core, it indicates yalmip('clear') has been issued since variable was created and user should be informed about the situation.

Formating of quadratic in savecplexlp

Clean trailing zeros etc.

kktqp crashes

Fails on everything

Support quadratic cones in outer approximations

Convexity propagation fail for nonconvex objective without constraint

Correctly detects non-convexity

sdpvar x
p = max(exp(x),-x);
sdpsettings('allownonconvex',0);
[F,failure,cause] = expandmodel([-1 <= x <= 1],-p,ops); 

But not here!

[F,failure,cause] = expandmodel([],-p,ops); 

Skip bilinearization if possible

Try to avoid to bilinearize models in bmibn if possible (nonlinear solver supports polynomials etc)

Domain dependent convexity propagation

Should be easy to do, the basis is already in there. sin is one example

Crash in problem compilation for indefinite quadratic

Error happens when the variable eae=0.2, but replace with 0.21 and it works. Seems to generate a quadratic with very small negative eigenvalues, thus triggering the code to treat marginally indefinite problems as positive semidefinite.

Note: Problem is nonconvex as it is bilinear.

Expected behaviour: No crash in either case, report as nonconvex?

Hessian fails on linear expression

To reproduce

sdpvar x
hessian(x)

saveampl does not expand complementary constraints

To reproduce
sdpvar x y
saveampl([complements(x>=0,z>=0),0<=x<=1,0<=z<=2],-x-z,'amplfail.mod');

Send Jacobian sparsity pattern to SNOPT

Generalize code from IPOPT interface

Cuts are not removed before picking upper bound solver for BMIBNB

Hence, if semidefinite cuts are involved, a upper bound solver such as fmincon will not be picked, even though the problem is a simple nonlinear program, without the cut

Reference on all options

Create some strcutured description of all non-external-solver specific options.

Support socps in nonlinear solvers

Automatically cast SOCPs (which arise if you use norms) as quadratic constraints if the solver supports this.

Kronecker slow

Trying to do

A = sdpvar(8760/4,40)
B = kron(A, ones(4,1))

is very slow, compared to

A = randn(8760/4,40)
B = kron(A, ones(4,1))

Robustify fails on quadratic wth uncertain hessian

x = sdpvar(2,1);
a = sdpvar
solvesdp([x'*(eye(2)*(1+a))*x <= a,uncertain(a),-.1 <= a <= .1],sum(x))

Cut operator for SOCPs

Does not work if SOCP originates in a high-level model, since only the lifted constraint is marked as a cut. The internally generated SOCP will be a normal constraint

Remove silly constraints

This causes a bug
[0<=1, x>=0]

This doesn't!
[x>=0,0<=1]

Wrong termination code when ooqp succeeds

ooqp reports success, but YALMIP reports time limit exceeded

setdependence slow

yalmip('clear')
N = 5000;
w = sdpvar(N, 1);
previousPos = rand(N, 1);
constraint = [0.0001 <= abs(w) <= 0.04, sum(abs(w)) == 1, abs(sum(w)) <= 0.04, sum(abs(w - previousPos)) <= 0.1];

Complex exponential

Reformulate to cos/sin

callipopt crashes if savesolveroutput is turned on

lmi/plot should not return anything if not asked to

sdpvar x
plot([ x <= 0; x >= 1])
ans = 

    {[]}

You should put something like this when exiting lmi/plot:
if nargout==0
clear x_opt
end

Options for SNOPT

Add possibility to control options in snopt

Too long lines in savecplexlp

Some solvers are strict with 255 chars per line, so code should break up long expressions.

modelComplementarityConstraints not synced with expandmodel

Variable bounds are extracted twice from model, e.g., in saveampl and compileinterfacedata

Individual test-code for every solver

Attach solver definitions with test-scripts to be run to detect issues etc. A field is already there, just has to be populated, and some layer around it to make it useful

Optimizer misses trivially infeasible model

Models which instantiates to trivially infeasible SOCP cones are not handled correctly

yalmip('clear')
sdpvar x
sdpvar z
sdpvar y
P = optimizer([-1 <= y <= 1,y >= z*z],x^2,[],[y;z],x)
[~,err] = P{[0;2]};

Account for unused variables in polytope conversion

Support the format polytope(P,x)

mtimes fails

To reproduce

yalmip('clear')
sdpvar x y z
x = x^2;
y = y^2;
z = z^2;

p1 = (x+y)^4;
p2 = -8_x^2;
(x+y)_(x+y)

Clean code wrt qcqp->socp->qcqp

Reverse the login in compileinterfacedata when checking for which solver this not should be done. reverese to for which it should be done

Unbounded problem in quadprog listed as infeasible

clp crashes on QP without constraints

Dualize fails

clear all
clc
X = sdpvar(4,4);
% Imposing psd on full X: Works fine.
% Imposing psd on a few submatrices only
submat1 = [1 2 3]; submat2 = [1 2 4];
F = [X(submat1,submat1) >= 0];
F = [F, X(submat2,submat2) >= 0];
% Some random additional constraints
F = [F, X(1,2) == 2];
obj = trace(X);
ops = sdpsettings('dualize',1);
solvesdp(F,obj,ops);

Matrix slog fails

log2(1+sdpvar(2))
Error using slog (line 23)
SLOG only defined for vector arguments

Constrained SOS

Add an example to show how to work with multipliers etc

Loading of sdpvar objects renders YALMIP unusable

After an sdpvar is loaded from disk, YALMIP stops to function properly.

To reproduce (Matlab R2011a, YALMIP e227e5):
sdpvar x
save x
clear
load x
clear
sdpvar y
double(y)
??? Cell contents reference from a non-cell array object.

Error in ==> yalmip at 617
varargout{1} = internal_sdpvarstate.optSolution{internal_sdpvarstate.activeSolution};

Error in ==> sdpvar.double at 37
solution = yalmip('getsolution');

Plotting of non-convex sets

It would be uber-cool if YALMIP could plot non-convex sets. Admittedly, this is more a wish rather than a reasonable request.

sdpvar x1 x2
plot([ x1^2 + x2^2 >= 1 ; -2 <= [x1; x2] <= 2 ])

Optimizer claims infeasible incorrectly

When variables are eliminated feasibility is incorrectly determined

sdpvar x
sdpvar z
sdpvar y
P = optimizer([-1 <= y <= 1,x <= 10+z*x],x^2,[],[y;z],x)
[~,err] = P{[0;5]};
err

The reason is that the ordering is messed up. Internally somewhere, it is assumed the parametric variables are listed in the same order as they were defined

Plotting of 1D sets does nothing

sdpvar x
plot([-1 <= x <= 0])

Actual behavior (with e227e): no figure produced
Expected behavior: plot the 1D set

Options missing for gurobi

BarHomogeneous for instance is not available

Convenient generation of envelopes etc

Add GEVP solver

Implement internal GEVP solver, automatically extracting the GEVP model from a bilinear SDP