|
From: | Jonathan Kinsey |
Subject: | Re: [Bug-gnubg] Multi-threaded cubeful #Q0 bug |
Date: | Wed, 28 Jan 2009 13:34:57 +0000 |
Christian Anthon wrote: > Jonathan Kinsey wrote: >> I'm pretty sure this is a bug in the cache code that doesn't treat the >> cubeful >> equities correctly. I'm in the midst of changing the cache code so >> can't check >> in a fix at the moment. The fact that is seems to work on some machines is >> probably a mixture of the default values being copied (0 vs rubbish) >> and the >> amount of cache usage (number of threads). >> > If that is the case turning of the cache should cure the problem. I > believe that you can do "set cache 0" in the command panel to turn off > the cache. Could somebody who observe the problem try that, please. I just tried that (I can reproduce the problem with the latest install but not if i build the code) and it worked fine after "set cache 0". >> I think the fix is to set the cubeful equity to zero when adding it to >> the cache >> (or maybe copy it if this is calculated?). The fix might be just adding: >> ec.ar[ OUTPUT_CUBEFUL_EQUITY ] = 0; >> before the CacheAdd calls (or in the CacheAdd function). It looks like >> the 6th >> value isn't set either (equity?) so this might need fixing as well? >> > I don't know too much about how the cache works. However, plied cubeful > evaluation are not a function of the win/g/bg probabilities alone. So I > think that both the cubeful and cubeless equties (the 6'th value) should > be right. Yes but only for cubeful analysis right? I think most of the analysis e.g. EvaluatePositionFull() only returns the 5 w/l stats. Jon Windows Live Hotmail just got better. Find out more! |
[Prev in Thread] | Current Thread | [Next in Thread] |