If Houdini goes to KRBBKRN then it is obvious it does not understand the position, Houdini 3 thinks it is a draw. So the only way to show that the new Houdini does not do this would be to repeat the experiment with the new version. Stockish does not really have a lot of endgame knowledge here, it just goes by the material and more specifically the material imbalance table. But if Houdini steers towards a lost endgame anything is possible. In the case of Rainbow serpent it also does not have extra knowledge and of course also no tablebases. Material imbalance table and such is slightly different. The interesting thing for me is if on longer searches it would avoid trading rooks, what it did in the short shootout, and maybe then White is safe. There is a very slight bonus for having a rook if you have a passed pawn, but it is not tuned and the opponent rook makes the bonus less. There is a bonus for rook behind the pawn but that also does not make much sense if the pawn can't march. In combination with the fact that the bishop pair will get a little more after the rook exchange, the high score for the initial position, no tablebase probes, short timecontrol etc, then the bonus for keeping the rook may not help significantly. But a Houdini 30 ply search is not really a proof.Laskos wrote: I am not sure, I played two games at long TC with Houdini 3 being white and Stockfish 2.3.1 black, Houdini lost both in the same fashion. There is a long maneuvering which I do not understand out of which, in 20 or so moves, Stockfish takes the pawn clean, then it's RBB against RN, which is won. Even if it's really a lost position for white, it might be a special case hard to put as a rule.
Kai
Regards, Eelco