On 16 January 2013 11:11, Philippe Michel <address@hidden>
Supervised training. I used the same training tools that were used years ago to create the current nets.
On Mon, 14 Jan 2013, Mark Higgins wrote:
What training approach have you been using, if you don't mind elaborating?
The main difference is that I rolled out the training database while it previously used (as far as I know) 2ply evaluations from the preceding generation of nets.
This obviously took some time, but with current processors what was out of question in the early- to mid-2000s when the currents nets were trained is now doable.
I don't know if Joseph Heled did many iterations (reevaluate database / train nets / maybe add mishandled positions) but with rollouts, each of them take a long time (I did it twice for the crashed database and once for the contact one). This is then mostly a one-shot effort, at least until something important changes in the training database.
Oh yes. Many iterations :) but at 2 ply, no rollouts.
I am willing to believe the new nets are better, but I have not seen the results of a long-enough/statistically-significant run of matches between the old and new.
Another thing that must have been helpful is that I added to the trainig databases its positions with the other player on roll. I think this helped a little for the general playing strength and diminished significantly the odd/even plies discrepancies.
I used slightly larger pruning nets, with sizes adapted to SSE or AVX instructions, but I don't think it make much of a difference.