The problem is similar to the "trolling" behavior of Lc0. Versions since 11258 have been trained with resignation on, presumably in positions less bad than knight odds. So it has no understanding of how it is worse to be down two knights than one (just for example). The PVs often include ridiculous moves, even blunders that a 1200 player would be ashamed of. As long as the initial eval of the losing side (as a winning percentage) is double digit or at least close to that, everything seems to be fine, But with win probabilities like 2 or 3 % it outputs nonsense. Version 11248 or 11258 gives double digit win prob for knight odds so it is ok at that handicap.mmt wrote: ↑Thu Sep 12, 2019 7:26 pmInteresting. Is the reason you're saying it's ridiculous for b1 handicap is because of a very negative evaluation? That's what seems to be happening at the beginning of the analysis but then the evals become somewhat less negative. I'll run it for longer (net 42850). A few games vs other top engines would also be interesting.
Opening book for game with handicap
Moderators: hgm, Rebel, chrisw
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: Opening book for game with handicap
Komodo rules!
-
- Posts: 343
- Joined: Sun Aug 25, 2019 8:33 am
- Full name: .
Re: Opening book for game with handicap
Thanks for the explanation!