| |
BGonline.org Forums
Non-doublet Replies: Five rules for beginners
Posted By: nack ballard In Response To: Non-doublet Replies: Five rules for beginners (Axel Reichert)
Date: Wednesday, 19 April 2017, at 12:40 a.m.
Good job, Alex -- your work ethic was better than mine. I took shortcuts and let a few positions slip through the cracks.
I have a few general comments:
In previous post I wrote "no mixing" under bot data. I believe that (actually, based on my past verifications) comparing a rollout equity with an evaluation equity (or differing evaluation levels with each other) produce larger errors in margin. If there is a play in question for which you have no rollout, it is better to evaluate ALL plays on the same level (++ preferred) and for the non-rolled out plays use THAT data to determine their margin.
Our best data is rollouts at the highest ply possible, as long as we have at least 5k. (I pretty much ignore 1k or less, except as a quick verification that some play is not best, and that is not what we're doing.) I'll assume for the time being that 2-ply checker / 3-ply cube is better than XGR++ evaluation, though I've never tried to compare both against a higher-ply rollout.
I think it is worthwhile pondering the tradeoff between (A) saving a rule, and (B) either meeting a round-number threshold, or avoid having to list a specific exception. For example, suppose you can save a rule by knocking back the .039 threshold to .041 or .042, then I think the weaker threshold is preferable. Or suppose you happen to devise a scheme whereby your threshold is .058 except for that pesky 32Z-64P (-.068) position plus one other. In that case, I think it's a better bang-for-the-buck to list those two positions (perhaps even three) outside the rules as exceptional.
Counting the number of rules is subjective. The more complex a rule, the more reason there is to divide it up into more than one rule. In some cases, using a corollary or convention (especially if it is repeatable) can be helpful for minimizing the number of rules.
Below is a table for the plays that has concerned one or both of us, though there are and will be others. The left column lists the position. The second column is the 3-ply rollout (that included the weak play in five of the thirteen cases that I could find). The third column is the 2-ply rollout, and the final column is the XGR++ evaluation. (The weak play that concerns us is the last one within the nacbracs.)
Posit...3-ply Rollout...2-ply Rollout.........++ Eval
32D-43...................................... [D Z19 S94]
54D-41 [$ W57 S81]"<5, [$ W49 S74]":<10,.....[$ W53 S56]e
43D-64...................................... [R S2 P77]e
64S-65 [K H75]"<=5,....[K D62]":<10,.........[K D56]e
51$-64 ..................................... [R Z64 U72]e
54S-61 [P H46 X64]"<=10*5, ..................[P H53 X80]e
52S-51 ..................................... [S U36 X63]e
41$-64................ [R Z60 U63]":<10, ....[R U50 Z54]e
52S-63 ..................................... [R S26 H31 X61]e
63S-65 ..................................... [K D61]e
52D-43 [D Z31 S59]"<=31*5, [D Z31 S58]":<10,.[D Z25 S55]e
32Z-64 [H R36 S45 P59]"<15*10, ..............[H R42 S45 P68]e
62S-65................ [K D56]":<10, ........[K D52]eFor replies to my 27-opener set (same as Jeremy's, minus 62R, 62$ and 43U), I propose this modified set of five rules that yields a threshold of under .06, and significantly reduces or erases previous errors:
...1) Hit >4pt
...2) Point or 65R (exc 7*1*)
...3) Hit on 4pt, or on 3pt with non-5
...4) Versus D, run past or x1$
...5) 43Z or S (if legal, else D)When you hit (i.e., #1 and #3), if leftover die, use the "CRMU" convention (Cover, Run, Mid, Up).
[The largest error (-.059) seems to be 32Z-64P, but opening 32Z is seldom seen. Did I miss anything? What is second largest?]
Nack
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.