| |
BGonline.org Forums
response to Bob Koca
Posted By: Jim Stutz In Response To: small clarification (Bob Koca)
Date: Sunday, 15 January 2017, at 5:28 a.m.
Bob Koca wrote: “So the expert would have to lose more than .28 to make it a take and if that is likely then you aren't playing an expert.”
Maybe this particular position wouldn’t qualify as sufficiently complex for the attacker, but I believe that thematically similar positions routinely arise from which an expert attacker will indeed, on average, lose .14 more equity playing his checkers during the rest of the game than his expert defending opponent.
Is there any contemporary data out there to support, or refute, this argument, and the more general argument that playing GG is more complex and mistake-prone for humans than GS? I’m sure there is -- thousands of matches have been transcribed and analyzed by XG -- but I don’t have access to this database. What I do have is the 25-year-old Woolsey MET, which was based in large part on gammon rates and other results taken from 1000+ matches (a small sample, but still large enough to have some validity, IMO) between the best human players of the era. I believe that the human-vs.-human data underlying the Woolsey MET, ancient and otherwise flawed as this data is, still provides some meaningful support for the argument that gammon prosecution is more mistake-prone for human players than GS.
Take the simplest example: At 2a/Crawford the Woolsey MET gives the trailer a 30% chance to win the match, whereas the corresponding value in the R/K MET is 32.26%. Doing the math, and allowing that GG will lose a few extra games while taking some chances to win the gammon he needs, the Woolsey MET implies that the 1990’s human trailer will win about 220 gammons for every 1000 games he wins against his GS human opponent ATS. Meanwhile, the R/K MET tells us that the bot trailer is winning about 265 gammons out of 1000 wins ATS -- more than 20% more gammons than his 1990’s human counterpart. This is true even though the bot has an equally talented bot as its opponent on defense, using all of its considerable skill to avoid the fatal gammon. This means that the best human players in the 1990’s, when on the GG side at 2a/Crawford, on average were donking off a double-whopper with cheese (~.226) more in checker-play equity per game than the GS human across the table. This works out to a checker-play ELO advantage of ~200+ for GS over GG, assuming an average game length of around 25 moves per side. Admittedly, open players were not as efficient at winning gammons in the 1990’s as they are today, but presumably they weren’t as skillful at defending on the GS side either. Despite the absence of contemporary data, I believe this retrograde analysis, tortured that it may seem (Ya think? Lol!), offers some support for all of the following related arguments:
1) On average, gammon prosecution presents more complex play and more opportunities for checker-play error on the attacking side.
2) In positions that are thematically characterized by asymmetrical complexity, such as GG vs. GS, bot-vs.bot-generated cube analysis becomes substantially inaccurate as a tool to guide humans to the best cube action.
3) The weaker your opponent, the deeper (relative to XG’s advice) you should take his cube if he is likely to face more complex checker play than you as the game unfolds.
4) Certain important values in the XG-generated R/K MET, particularly those at scores close to the end of a match when gammons have value only to the trailer, are significantly biased by XG’s superiority over humans in gammoning skill and efficiency.
Stick wrote in an earlier post that disagrees with most these arguments. Obviously he knows a great many things that I don't, so I'm looking forward to his response when he gets around to posting it.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.