First of all, I want to thank all those who realized that there were problems with the QP solver of MATLAB. They simplified my life

. After using the noise trick for getting an alpha_0 and putting an upper bound of 1e6 my solvers work

.

Just want to clarify what is meant by "How often is gSVM better than gPLA in approximating f". I understand this as, per each run, we estimate E_out for both the PLA and SVM. If the SVM has a lower E_out than the PLA, then we count that run. After performing the 1000 experiments, we divide the number of runs counted by 1000 and multiply by 100. Correct?