| |
BGonline.org Forums
XGR++ RO and 1 ply RO..etc
Posted By: Henrik Bukkjaer In Response To: XGR++ RO and 1 ply RO..etc (Daniel Murphy)
Date: Thursday, 15 December 2011, at 1:16 p.m.
Daniel, I agree a lot with what you say. (Stick mucked gnubg, not just the "noise" part, but that's beside the point).
Anyway, do noise do any good?
It really do produce "random errors", that in no way mimic certain playing styles or human error patterns. We tend to see some position types as easy and some as difficult. Gnubg noise doesn't see things that way. In fact, what it finds "difficult", are positions where a lot of moves are within a relatively small error span (making it more likely that noise will pick an alternative with less equity). In fact, adding just a little noise will never result in huge errors, but in could result in "obvious" errors from a human perspective.
I've been experimenting with this, and found, that the noise typically add more errors late in the game, where humans ofter play with a lower error rate. This is due to more alternatives close together in equity, in bear-ins, etc. That's a problem for sure.
Does it matter for this purpose, which errors it produce anyway? I'm not sure. For some purposes it's hopeless, such as comparing two positions or sides, and see which one is most affected by noise to determine which one is more difficult to play... Doesn't work at all...
But here, where all you want to do, is to simulate, that one player will not win his fair share of the games played from a given position? I know you have to estimate avg losses for that player, and then adjust noise accordingly, so you get the same net result.
Eg.: From here on, optimal bot-play will give 32% wins. My guess, over the board, would be that this player will err on avg to get only 30%. So I add just a little noise, and verify the result of that. By doing so, over a long period of time on many positions, you'll get a good "feeling" for the effect of noise.
And it simply must be better to perform your rollouts with error-patterns that are off in the distribution, but "on" in the resulting win percentages, compared to performing a rollout with a setting that will give wrong win percentages!
What problems will the wrong error-distribution yield then? Well, I suppose it will give some problems when the cube is available to the player, if the rollout is performed with live cube and play accordingly. Because the equity-loss would be a little delayed in the actual game, making mid-game cube situations happen when the player is a little bit better into this game, but after the cube action, he will then start loosing more equity than the evaluations thought he would. So, for cubeful purposes, it suck. Especially since you cannot add the noise "one-sided", each nn-eval will not be skewed in the same direction as overall play. It should be so, that when you add noise to checkerplay, all cube decisions should have the noise added to subtract only in the evaluations (making the resulting noise ditribute evenly around the actual equity, due to expected bad checkerplay for the remainder of the game.
This problem - I guess - would result in one thing, if you play vs. gnubg with noise and evaluate afterwards: Typically worse cube-actions than checkerplay (remember, the cube-actions should be considered vs. the Jac50 or Jac100 table.
Now, even better would be, if you could add noise differently to the nets used, so that you could decrease the noise added to non-contact positions, etc. And also, if you could add noise unevenly distributed around the true equity (eg. +0.005 -0.015) so you avoid the early double syndrome. That would be improvements, while not perfect.
Best regards, Henrik
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.