| |
BGonline.org Forums
Example of recently discussed XG bug
Posted By: Timothy Chow
Date: Tuesday, 27 October 2015, at 11:06 p.m.
In a recent thread, a potential XG bug was discussed at some length.
You can download a match file of mine where I experienced a similar bug. If you go to the last game in the match and look at the last decision in the game, you'll see that it's flagged as a blunder by XGR++. But that's because XGR++ has an absurd evaluation of its favored play.
One problem with this example is that I haven't been able to figure out how to reproduce it. If you click on the offending play and ask for an XGR++ evaluation of it, it will re-evaluate to something reasonable.
I gather that one theory for what happened here is that it was some kind of race condition or operating system problem, associated with running other things on the computer. I have doubts about that theory, for a couple of reasons. First, I don't run other things on my computer, other than the normal background processes that are always running, such as the antivirus software. The computer is not even connected to the Internet. Second, the problem didn't occur just once. I play matches in competition mode, and I also turn off XG's real-time evaluation function because I'd rather not see what my estimated PR is for each game before I proceed to the next game in the match (though I haven't figured out how to disable this feature entirely, since it will tell me whether I made take/pass errors). Then I ask XG to analyze the match on World Class level. In this case, my move didn't make it through the initial move filter, so it was analyzed on XGR+, which charged me with a blunder. Incredulous, I manually forced an XGR++ evaluation of this particular decision after the match analysis was complete. The results are what you observe in this file. If it were a random race condition, then it would be highly improbable for the race condition to occur twice like this, first for XGR+, then for XGR++.
Reproducing the problem might require me to play the match again from scratch in order to get XG into the same state. There should be enough dice rolls in the match to reconstruct the random number generator seed, although even then, perhaps manually setting the seed will get XG into a different state from the state that originally caused the problem.
Note, by the way, that if you copy and paste the position into a separate copy of XG and ask for an XGR++ evaluation, you'll get a slightly different value. This seems to confirm that XGR++ depends on not just on the position but on the fact that the position is embedded in a match file.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.