| |
BGonline.org Forums
Choose chicken! Choose beef!
Posted By: Sam Pottle In Response To: BOT COMPARISON 400 / 500 GAMES (Joe Russell)
Date: Tuesday, 30 March 2010, at 10:07 a.m.
Once you have identified an error you have also identified the type of error
Have you? Suppose you have a position that 4-ply says is a double and XGR+ says is a no-double. Who should you believe?
FOUR: Don't believe Plus, she misses doubles all the time!
PLUS: Don't believe Four, he makes an awful lot of wrong doubles!
The reverse case presumably comes up less often, but each side can again present a reasonable argument.
There's also an economic problem: do you want to do an XGR+ eval of a position in order to find out whether it's worthwhile to do an XGR+ eval of the position?
There is a sense in which the type of error is determined by the play made OTB: if you doubled, either you were right or you made a wrong double, and if you didn't double, either you were right or you missed a double. I suppose you could use 4-ply to evaluate your no-doubles and XGR+ to evaluate your doubles, since each is less likely to wrongly agree with your play. But do you really want to select the arbiter of the truth of a position based on the decision taken at the table? This seems like a biased process. At a minimum, I would think, you should be willing to roll out all the decisions that are marked wrong by this process, since there will be more false positives.
Here's something you could do, as long as you're willing to run XGR+ on all your cube decisions: evaluate each decision using both 4-ply and XGR+. Roll out every position on which they disagree.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.