On Fri, Apr 26, 2024 at 06:04:50PM -0600, Murat K wrote: > I posted an article including pretty images of some positions on: > > https://www.bgonline.org/forums/webbbs_config.pl?read=213189 > > Please read it there and then feel free to post replies here if > you prefer also. It may not be a bug by definition but it's about > GnuBG, involving a very strange game that lasted for hours and > who knows for how many hundreds (if not thousands) of rolls...
I have seen somewhat similar positions in the training database. Maybe not as extreme but ludicrous stacks in the outfield. As far as I know, this database was created from bot vs. bot play too. I looked for these positions and added slightly derived ones, for instance if there was a 5 checkers stack I added one with a 131 pattern centered on the same point, but since the whole positions were very unnatural it is not even certain that the the smoother position would evaluate better. Even if it does, tipping the neural network enough for it to dislike these positions with just a few such additions was rather unlikely. Something more radical may have been needed. In supervised learning, would it make sense to cull the, say, 1% of the training set that is the more badly evaluated by a generation of the neural net before training the next one ? On one hand they may be impossible with sensible play and somehow poisonous, on the other hand they may be useful but difficult. Since it amounts to thousands of positions, looking at every one of them to decide would be hard work.
