| |
BGonline.org Forums
New GNU
Posted By: Philippe Michel In Response To: New GNU (Michael Depreli)
Date: Thursday, 4 April 2013, at 9:53 p.m.
The last significant change related to speed dates back from July 2011. If your 0.90 version is more recent than that, any difference should be minimal.
Regarding your second question, there were two problems with odd/even plys :
- evaluations could swing wildly from one ply to another : I think this is significantly alleviated. In many cases equities stay very close from 2 to 3 to 4 ply and the remaining variations are generally smaller than they used to be (0 to 1 to 2 ply doesn't seem as reliable).
- odd plys were barely, if at all, stronger than the preceding even ply. Benchmarks used during training to select the candidate nets showed that, on average, 2 ply is clearly better that 1 ply which is clearly better than 0 ply (except for races in the latter case : 1 ply is only about as strong as 0 ply). I think this is the case for higher plies as well and that 3 ply is now better that 2.
This suggest a possible speed improvement, especially for 2 ply rollouts :
Currently, default move filtering is "evaluate at 0 ply then evaluate at 2 ply up to 8 moves within 0.16 of the best 0 ply choice". There are narrower move filters but I doubt they are much used
Using something like "0 ply, then up to 8 within 0.16 at 1 ply then up to 4 within 0.08 at 2 ply" should be nicely faster and still accurate for most positions.
This is probably a little reckless for match analysis (if there are many close choices, I'd rather have them all evaluated at the top ply), but for rollouts it could be good enough.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.