Adjustment for GNU errors
Posted By: Nack Ballard In Response To: Adjustment for GNU errors (David Rockwell)
Date: Sunday, 3 January 2010, at 12:35 a.m.
In Response To: Adjustment for GNU errors (David Rockwell)
First, let’s get the mundane stuff out of the way. You asked to be informed of corrections. This may not technically qualify, but for 63R-31 and 63R-63, the character of X (for hit and split) is standard rather than S (though the latter is unambiguous, and I won’t complain if S is your conscious preference). I suspect you know this but are subconsciously confusing situations like 21$-43S where you can’t use X because the hit and split is made with the same half of the roll.
For the header above the nacbracs, I suggest something like “63R at GS, errors at GG” or “63R-xx, errors at GG” or “63R at GS: reply errors” instead of just “63R.” IMO, it is better that the explanation in the paragraph preceding the table supplement rather than replace clear header or entry information. (The reason I emboldened 63R here is to emphasize contrast with 63S in the other heading; just an optional frill.)
Rather than elucidate the 63R and 63S headers, I chose to do away with them and insert “g” (nacbrac convention for gammon go) before the brackets because there are only a few entries and I will need to reference a position in my text where GG clarification is necessary.
I’m repeating your data below, with my own formatting. Yours was fine, but I like to save characters (to minimize eye-scrolling) when I don’t feel readability is sacrificed (more than an iota). To that end, I underlined Gnu’s misplay rather than repeating it to the left of the brackets. (To avoid the work of creating html tags, another choice would have been to state, “In each position, Gnu makes the second play listed.”)
63R-31 g[X H18] 5k .. 63R-51 g[$ S24] 11k .. 63R-54 g[X D3.3] 15k .. 63R-63 g[D X13] 10k .. 63R-65 g[H S12] 5k
63S-32 g[$ H11] 5k .. 63S-51 g[H K3.7] 15k .. 63S-62 g[X D0.1] 15k
The reliability of some Gnu money reply rollouts is in question. Given an adequate number of trials, I assume/estimate (rightly or wrongly) that half of whatever bias exists in its money margin will carry over to the GG result.
The only real bandit of this group seems to be 63R-51 g[$ S24]. For money, Gnu is [S $4], Snowie is [S $22] 31k, and XG is [S $13] 36k, for a bot average of [$ $13]. From this, I gather that Gnu has a money rollout slot bias of .009. If half of this bias exists in the GG rollout, that comes to .0045. So, my adjusted value is 63R-51 g[$ S19.5].
For some positions there is no Gnu money rollout (due to lopsidedness), so I didn't have the information needed to compute all carryover adjustments. In the end, I made (or left) the 63R/xx GG errors 19.5, 19, 3.3, 13, 12, and the 63S/xx GG errors 10.7, 4.5 and 0.0.
Summing, dividing by 18 and putting the decimal point back where it belongs, this comes to an overall rollout adjustment of 63R = .0037 , and 63S = .0008, a difference of .0029. It turns out these differ very little from your numbers (not adjusted for money bias carryovers), which yield a difference of .0031.
Finally, you used your 63 s[R S5.4] 15k rollout as a base from which to adjust. I used a combination of rollouts (including yours), that I’ll call a base of s[R S4.6]. As my adjustment is .0002 less than yours but you’re using a base .0008 higher, there is a partial offset. I finished with [R S1.7] compared to your [R S2.3].
History of Opening 63 at GS: R was perceived to be better by .006 according to Jellyfish rollouts, and then confirmed (margin of .005) by Snowie rollouts, but checker play was not according to score. Along came Gnu, a long rollout was assigned and an overturn was dealt: now S came out on top by around .009 (as I recall) after 20k. At my request, Stick extended it to 31k, narrowing the S-preferred margin to .005, and posted it on his site. Several months ago, you and I independently deduced that this file is corrupt (though I believe it’s only the first 20k, not the last 10k). Stick was leaving for Europe around that time and never got around to pulling the corrupt 63R at GS rollout off his site, and there it still sits. We now have uncorrupted Gnu data -- in large part yours -- showing R to be .005 better (coming full circle back to Jellyfish and Snowie), but that’s taking the rollouts at face value. As best we can tell today, after adjusting for Gnu checker reply errors, R is only about .002 better than S, theoretically speaking.
Good work, David (and thanks to Neil for the suggestion).
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.