Quote:
Originally Posted by Anne Paulson
No, the algorithm will always converge eventually if there is linear separation, no matter how you pick the misclassified point. And the proof isn't that difficult.
But now we have one person in one thread saying that they found picking the points at random gave better results, and another person in another thread saying picking the first misclassified point gave better results. Mysterious.
My strategy would be to start with the last misclassified point, the one I just updated with; try it again if it is still misclassified, and then pick the next misclassified point in the list.

If you are referring to my post then I am sorry, it was a typo, randomization is faster always. About your approach, I think it is necessary that the last point is always classified correctly due to the update rule, isn't it?