| |
BGonline.org Forums
Agg calculation
Posted By: Nack Ballard In Response To: 4-ply queue (Matt Ryder)
Date: Thursday, 6 January 2011, at 6:44 a.m.
By the looks of it, this ambitious replies project is nearing completion! Is there any chance you might post a consolidated list of results at that point?
That would involve some work, some of which would be redoing what I've already posted. Perhaps it would be best if someone searches and makes a list of the links for each 4-ply rollout post. Or perhaps even better: make a list of the links where I responded to each post, listing the data after someone posted an XG 4-ply rollout; that way the results for all bots will be there, often along with the agg and some related comparison positions.
If you (or whoever) is interested in doing that, you might want to first compare notes with Kye Hedlund. He might already have some of the links or you could divide up the search work.
Also, what's your methodology for deriving a bot aggregate (agg)?
If every play is at least 15k trials, I use (Snowie + GnuBG + (XG 3-ply + XG 4-ply) / 2) / 3, then round towards XG (4-ply, unless its fewer than about 25k trials and the number of 3-ply trials dwarfs it).
For plays with fewer than 15k trials, I tend to use a weighted average of trials, but cap each bot's input at 25k (and still treat all 15k+ results as equal); for example...
41$-33 [E C7 B17] ~20*10, [C E6 B25] /15*5, [E C1 B15] <145*103, [C E4 B25] ^46*5
C over E is (-7 + 6 + (-1+4)/2) / 3 = 0.17, rounded to XG 4-ply is [C E1]a.
C over B is ((10x10) + (25x5) + (16.67 x 25) / 40 becomes [C E1 B16]a.(Note that when figuring C over B, E is dropped, so that the first and third results are reduced to [C B10] ~10 and [C B14] <25 (cap), respectively.)
This algorithm can probably be improved on, but I wanted something that weighs the bots input equally (shading towards XG) yet takes variance into account. A pure weighting of trials wouldn't account for the difference in bot approaches; hence the 25k cap compromise. I also wanted something that isn't overly complicated, as I've been doing it manually (usually in my head). In truth, having done it so many times, I typically just eyeball it and get the same result (when I've checked) within 0 or 1 thousandth.
If you have another idea for determining agg, I'm open to hearing it.
In case some people have a radically different idea than I do about value (e.g., "I'm worshipping XG 4-ply's result even if it's only 5k and ignoring the rest"), more often than not I itemize all bots' results even if I also summarize the position with an agg.
Nack
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.