[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Wild rollout results (bug?)

Posted By: Nack Ballard
Date: Thursday, 22 April 2010, at 1:05 a.m.

In Response To: 3 ply RO of 21$-54..over 100k (neilkaz)


2O ' ' '1X4X '3X ' '1X5O

2X ' ' ' '5O '3O ' ' '4X

21$-54


Thanks, Neil. So, now we have...

Snowie .. [S R12] 10k
GnuBG .. [S R3] 46k
XG 3-ply [S R13] 103k, and [S R40] 10k, [S R38] 10k
XG 4-ply [S R7] 19k

The question remains: "What is going on with Chase's two rollout results in bold?" They can be viewed here and here. (In case it matters, the first one is no Jacoby, and he attempted to use the same seed in both.)

I mentioned some (obvious) possibilities here, but maybe you or Xavier can see something amiss in the file output or someone else has an idea.

If Chase's results turn out to be extreme variance, then they're legit and his [S R39] 20k should be added to your [S R13] 103k, to bring the XG 3-ply total for 21$-54 up to [S R17+] 124k. Or if XG seeds are reproducible (I'm not sure I ever got an answer on that), then we count Chase's result as 10k and the total is [S R15+] 114k. You rolled so many trials (103k) that you managed to minimize the effect of his result on the average, but it's still good to be be aware that such wild results can and will occur.

OTOH, if Chase's two files combining to [S R39] 20k are bogus, the margin differs from that of your long result by .026. Does it matter, since we can conclude S is better anyway? For sure! What if you were rolling out a position and got play A beating play B by .013 when actually it is B that is better by .013? Or, worse, the rollout says A is better by .01 but actually B is better by .08? (That is, we don't know the nature or extent of the corruption or when it might strike.) You then misplay that position whenever it comes up and it throws off your intuition for all similar positions.

In other words, while I don't want to be an alarmist, if there is a bug in the XG program (marvelous tool that it is) or particular settings to avoid or some other problem to circumvent in its usage, it's better for all of us to find out sooner rather than later. [I can say that with assertion after fake results have led me down wrong analytic paths, then double-rerolling and eventually discovering several dozen corrupt Gnu files and still finding them; the downside is a lot more than just the cumulatively wasted CPU time.]

Nack

Messages In This Thread

 

Post Response

Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:

 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.