Cryptic error messages due to solver failure

1

Hello,

I am trying to solve the following optimzation problem in cvx (image because I don’t know how to type that in mathtype)

image description

and I am having some trouble with it. I am getting not very descriptive errors. My code is the following:

cvx_begin quiet
disp('searching for 2-norm soft margin kernel matrix weights and param - eqn 45');
         
% optimization variables 
% TODO - can we make this dynamic?
variables u1 u2 u3 u4 u5 u6
variable t
variable lambda
variable v(train_size)
variable tau
        
% objective function
minimize( t );
         
% constraints
subject to
    K_prime = u1*kernel_guesses{1} + u2*kernel_guesses{2} + u3*kernel_guesses{3} + ... 
              u4*kernel_guesses{4} + u5*kernel_guesses{5} + u6*kernel_guesses{6};
    
    trace(K_prime + tau*eye(train_size+test_size)) == trace_val;
    K_prime == semidefinite(train_size+test_size);
    v >= 0;
    tau >= 0;
            
    % LMI
    K_tr_prime = 
u1*kernel_guesses{1}(1:train_size, 1:train_size) + ...
u2*kernel_guesses{2}(1:train_size, 1:train_size) + ...
u3*kernel_guesses{3}(1:train_size, 1:train_size) + ...
u4*kernel_guesses{4}(1:train_size, 1:train_size) + ...
u5*kernel_guesses{5}(1:train_size, 1:train_size) + ...
u6*kernel_guesses{6}(1:train_size, 1:train_size);
              
[
 (gram(K_tr_prime, train_labels) + tau*eye(train_size))        
 (ones(train_size,1) + v + lambda*train_labels);
 
 (ones(train_size,1) + v + lambda*train_labels)'
                    t
] == semidefinite(train_size+1);
    
cvx_end

And my errors are the following:

Error using eig
Input to EIG must not contain NaN or Inf.

Error in maxstep (line 64)
    mindxs = min(psdeig(reldx, K));

Error in wregion (line 99)
        maxt1 = min(maxstep(dx,xc,uxc,K),
        maxstep(dz,zc,uzc,K));

Error in sedumi (line 444)
    [xscl,yNxt,zscl,y0Nxt, w,relt,
    dxmdz,err, wr] = ...

Error in cvx_solve_sedumi (line 99)
[ xx, yy, info ] = sedumi( At, b, c, K, pars
);

Error in cvxprob/solve (line 383)
        [ x, y, status, tprec, iters ] =
        feval( sfunc, At, b, c, cones,
        quiet, prec );

Error in cvx_end (line 77)
    solve( prob );

Error in main_aud_cvx>classify (line 274)
cvx_end

Which I can’t really do much with. What exactly am I doing wrong? Is it how I’m using the u1-u6 variables?

2

This looks like a solver problem, actually. Try SDPT3 and see what happens. It also appears as if you are using an older version of CVX, but I don’t think that’s an issue—though the error message would have been a bit more clear in the newer version.

3

I also opened a bug report for this question to address the non informative error message and in the result ended up upgrading to v2.0 and my problem was solved.

Ticket: http://support.cvxr.com/support/tickets/20

4

Thanks, by any chance do you know if its possible to dynamically create u1-u6 and use them inside cvx?

5

I’m glad we could work it out! I’m going to change the title of this thread so that it is a bit more directly related to the problem you encountered.