Pocket Algorithm
I'm working on coding the pocket algorithm, however I feel I have hit an impasse. I'm not sure if it is a bug in my code or a flaw in the inherent structure of the pocket algorithm, but when testing the code on a known linearly separable data set, the algorithm often 'settles' on a nonoptimal weight vector (i.e. one where Ein != 0). I believe this is due to the fact that the algorithm does not allow for weights producing higher values of Ein even if those values may later lead to an even lower Ein.
So, what I'm asking is, should the pocket algorithm always be able to find a linear classifier that separates separable data?
