| |
|---|---|
| Opening 43 | |
This has been my thinking as well, though I think it might well extend to weaker players as well (as anyone is more likely to err on the immediate reply vs 43U than one of the other plays), and not only at dmp but also at gammon save.
I extended the Snowie rollout to 46k in part to get a better spectrum of bots' money data. Then, if some bots appear to have a bias for or against certain plays for money, some partial adjustment might need to be made to the AtS data as well. In this post, I'll apply my adjustment methodology.
Here is data for opening 43, grouped in order of money, dmp, gammon save, gammon go:
Snowie .. [D S3 Z5 U6] 46k
GnuBG .. [D Z5 S13 U23] 92k 77kSU
XG 3-ply [D Z6 S10 U14] 108k 62kSU
XG 4-ply [D Z5 S12 U27] 46k 23kS, but only 10k of U
GnuBG .. d[Z U2.3 D4.4 S5.1] 46k 28kS
XG 3-ply d[D S2 U2 Z4] 15k
GnuBG .. s[Z S2.9 U6.3 D10] 36k 20kD
XG 3-ply s[U Z6 D10 S11] 5k
GnuBG .. g[D Z21 S33] 10k
Key: D = Down (13/9 13/10), S = Split (24/20 13/10), Z = reverse split (24/21 13/9), U = Up (24/21 24/20). Error sizes are in thousandths.
A lower-case letter before the brackets denotes score: d = dmp, s = gammon save, g = gammon go; else assume money. After the brackets is the number of trials (in thousands, rounded down). For example, "s[U Z6 D10 S11] 5k" means that at gammon save, U is best, Z is -.006, D is -.010, and S is -.011, after 5000+ trials.
Averaging Snowie, Gnu and XG results for money yields [D Z5 S9 U15] if we give XG 4-ply's (mere) 10k result of U half weight.
Gnu's money result of [D Z5 S13 U23] suggests an anti-S bias of .004 and an anti-U bias of .008, compared to D and Z. If we carry half of those perceived biases to gammon save, that adjusts its result to s[Z S0.9 U2.3 D10]. If a third of the biases are carried to dmp, that adjusts to d[U Z0.4 S3.8 D4.8].
Applying a similar bias carryover analysis to XG 3-ply (which fits the money aggregate much better than Gnu does) means only slight adjustments: gammon save to s[U Z5 D9.5 S10] and dmp to d[D S1.7 U2.3 Z3.7].
AtS, the Gnu rollouts are longer than the XG rollouts. With that in mind, I'll favor Gnu over XG by 3:2 at dmp and by 2:1 at XG (where the latter bot has only 5k trials). This yields
d[U Z0.8 S2 D2]
s[U Z0.1 S2.4 D9.8]
That is my best current estimate of theoretical dmp and gammon save values. One has to adjust from there for style and opponent factors. Personally, I would adjust U favorably a bit more based on the chance of immediate reply errors due to unfamiliarity -- IMO that overshadows everything else.
Therefore, unless I see some clear counterevidence (e.g., a XG 4-ply rollout with a lot of trials favoring Z over U at DMP and GS, and/or perhaps analysis that Snowie is misplaying replies to 43U more than other bots), I'm a supporter of U (as shown below) at both the dmp and gammon save scores.
Nack
| |
|---|---|
| Blue played opening 43U | |
