| |
BGonline.org Forums
Will XG2 have a 'View Statistics' feature?
Posted By: Daniel Murphy In Response To: Will XG2 have a 'View Statistics' feature? (Timothy Chow)
Date: Saturday, 9 April 2011, at 12:07 a.m.
Your strong language took me by surprise, because from the earlier discussions to which I linked above, I didn't get the impression that you had concluded that bot-generated "cubeless" numbers were so worthless, so often misused, and so incomprehensibly derived. So now, I'm not understanding why they're so valueless, and what, exactly, you think would be better.
Take for example a simple position: money game cube action with two checkers on the six point versus two checkers on the six point. Do a cubeful rollout. Your bot will report a "cubeless" or "pseudocubeless" win percentage of 75.45%. For this simple position, that's exactly right. That win percentage is not truly cubeless, because at each double/pass the rollout stops and the bot's evaluation is folded into the total along with all the trials played to completion. But for this position the bot makes no errors in its evaluations -- and its evaluations are often very accurate in positions more complicated than this one.
Is knowing this position has a cubeless win percentage (CPW) of 75.45% useful? I think so. It tells us that if this were a last roll position, correct cube action would be double/pass by 0.45%. But knowing that, we might suspect that even with just 12 pips per side to go, there's probably enough vigorish to squeeze out a take. And our rollout tells us that is correct, since the cubeful equity is +0.961.
Earlier, you expressed a desire for an option for a bot not to report "cubeless" numbers that aren't, we all agree, really cubeless. That's fine, but you also seemed to express a preference that the default be not to report them. And I'm not understanding why one should have a better understanding of many positions, including this one, not knowing the "cubeless" win percentage.
Now, in this thread, you've asked that XG have a "view statistics" feature like Gnubg has. And you know I think that's often an interesting feature -- you've made good use of it in several discussions, and I believe it was a post of mine that alerted you to the existence of the feature in Gnubg. However, suppose we roll out the position 46,656 times with Gnubg and use "view statistics" to see exactly what happens in the rollout after double/take. We learn:
Percent 73.41% Blue (on roll) wins 2 points 11.51% White wins 2 points without redoubling. 15.08% White redoubles to 4: 10.30% Blue passes so White wins 2 points 4.78% Blue takes: 3.20% White wins 4 points 1.55% Blue redoubles to 8, and White passes, and 0.03% Blue redoubles to 8, and White takes, of which Blue won 7/46656 games and lost 2/46656 on the R/P of the 16 cube. Is that interesting? Yes, I think it's somewhat interesting. But it's awfully hard to parse, isn't it, and thank goodness we had no 32-cubes or gammons on the 2-, 4-, 8- and 16-cube level to absorb. The numbers specify White's cube use, but don't really seem to add all that much to our understanding of the position.
We also learn from "view statistics" that, in comparison to the cubeless or "cubeless" CPW of 75.45%, Blue actually won 74.97%. That's rather less interesting, I think, since the CPW and the cubeful equity already told us to expect Blue's actual wins to be somewhat lower, having doubled. That the actual WP is exactly 0.03% less than 75% seems not very interesting at all, since the bigger factor in reducing Blue's cubeful equity below +1.000 is not the WP but the number of games White wins on the 2-cube level with D/P plus White's net number of points won on the 4- and 8-cube levels.
So, what's "unreasonable," I think, is seemingly dismissing intelligent use of useful "cubeless" numbers as "teareading" and the process of deriving those numbers as "convoluted" since the way in which the bot calculates its "cubeless" numbers is actually quite simple.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.