Trust XG2 defaults
Posted By: neilkaz In Response To: Which MET to use in XG (Rick Janowski)
Date: Tuesday, 25 June 2013, at 6:51 p.m.
In Response To: Which MET to use in XG (Rick Janowski)
Kaz-XG2 is the default MET for XG2 because it was rolled out using XG2 for all scores in a 9 pt match. I feel that it is more accurate from a technical standpoint since each score was rolled out 100k times rather than 46k times (if I recall our work with GNU) resulting in a lower standard deviation, and since XG2 is clearly a stronger bot than GNU was then hopefully the stronger level of play results in getting to closer to reality if perfection were playing vs perfection. How that is applies when player A plays player B, or when Giant A plays Giant B is, of course unclear, but not the point of making these MET's.
XGR cube in XG2 is a tiny bit stronger than 4 ply cube so I used that, since it didn't add too much to overall RO time and since I want to minimize the effects of critical cube action errors in score based positions. I used XGR+ cube for 2 away PC since it did a better job of judging when it was a playon after a great opening roll sequence.
Therefore since Kaz-XG2 was rolled out with XG2, the strongest bot, and for at least twice as many trials and with clearly stronger cube settings than R-K, I'd use Kaz-XG2 as the MET for XG2.
Why did I stop after 9 pts when rolling out Kaz-XG2? I stopped because there was very little difference in the scores for matches of this length or slightly shorter between the two METs so, while I originally intended to go to at least 11 pts with my RO's I stopped and just used the R-K results for longer lengths.
Re: extrapolation for R-K from 16-25 pts. I spent a couple of days work on this starting with a calculated table that for many scores in the 11-15 pt range was very close to the R-K rollouts, but for lopsided scores needed some adjustments. I used my best mathematical judgements for those adjustments and and then checked take points to make sure that nothing stood as a big difference from what was expected or the expected progression of such. I've since tested a few of these match scores where the scores cross over from rolled out R-K to extrapolated R-K and am happy with my work.
We all should thank David Rockwell for starting the project to RO R-K with GNU 2 ply. This was necessary since bot RO's at GG using the initial ply (GNU 0, XG 1 ply) don't seem to be able to look ahead far enough in play to take best advantage of the fact that GV is 1 while GC is 0. I became aware of David's RO work and asked if I could help speed it up, having access to more and faster computers than he did.
I did not involve anyone else in creating Kaz-XG2 since I did it prior to public release of XG so it could be included as default with XG2. Yes, I waited until Xavier locked down the NN's for XG2 prior to release before starting the MET RO's so we can be certain that it was done with the same NN's we all enjoy from XG2 today.
If we do XG3 I'll probably RO a MET again for it and will again use 3 ply play XGR cube, but with something longer than 100k trials to start with. Perhaps I'll see nothing more than a 0.02% or so difference from Kaz-XG2 and give up, or perhaps not.
Another default I sometimes see questioned is our choice of 3 ply play XGR cube as the default for XG2 RO's. This is a good compromise between speed and strength. XGR cube is considerably stronger than 3 ply cube and while considerably slower, its use doesn't cause overall RO time to be greatly longer. We feel that this is worth the improvement from using 3 ply both.
.. neilkaz ..
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.