Learning feature

[Hood]

Hi,
the old programs as Rybka, Fritz etc. are capable to learn from a position analysis.
Are the new ones Houdini, Stockfish, Fire able to do the same. ?

Rgds
Hood

[kingliveson]

Hi,
the old programs as Rybka, Fritz etc. are capable to learn from a position analysis.
Are the new ones Houdini, Stockfish, Fire able to do the same. ?

Rgds
Hood

As for IvanHoe, it was a planned feature, but the project has stalled. Houdini and Fire do not have it, and I dont believe Stockfish does either.

[Uly]

Fritz never had learning feature, it only has BOOK learning, and, anyway, the GUI is doing it, so the GUI can provide book learning for all engines not just Fritz. But it's useless for analysis. And also, it's badly flawed and a reason the playchess engine room (where most people there use Fritz book learning) plays mostly the Sicilian defense and most of chess is "pruned".

Rybka only had learning (Persistent Hash) in Rybka 3, and it was so full of bugs that I'm one of the very few people that use it.

The people that reverse-engineered Rybka 3 didn't care about preserving her learning mechanism.

The only engine that has learning properly (one that is useful) is Shredder.

As for Stockfish, being open source anyone is welcome to implement learning for it, maybe some day...

EDIT - I just realized that the controversial engines are also open source (discarding Houdini) so I guess anyone can implement learning on them?

[ernest]

the old programs as Rybka, Fritz etc. are capable to learn from a position analysis.

Rybka, Fritz: NO!, where did you see that?
But Shredder, yes!

[Hood]

Rybka has preserve analysis option.

Concerning Fritz in Fritz interface going forward and back influenced the preferred variation.

[Hood]

Fritz never had learning feature, it only has BOOK learning, and, anyway, the GUI is doing it, so the GUI can provide book learning for all engines not just Fritz. But it's useless for analysis. And also, it's badly flawed and a reason the playchess engine room (where most people there use Fritz book learning) plays mostly the Sicilian defense and most of chess is "pruned".

Rybka only had learning (Persistent Hash) in Rybka 3, and it was so full of bugs that I'm one of the very few people that use it.

The people that reverse-engineered Rybka 3 didn't care about preserving her learning mechanism.

The only engine that has learning properly (one that is useful) is Shredder.

As for Stockfish, being open source anyone is welcome to implement learning for it, maybe some day...

EDIT - I just realized that the controversial engines are also open source (discarding Houdini) so I guess anyone can implement learning on them?

Rybka 3 has had preserve analysis as well . Does Rybka 4 have it ?
Rgds Hood

[tano-urayoan]

Rybka 3 has had preserve analysis as well

But preserve analysis is not learning.

[Hood]

In the context of the current position analysis i assume that it is learning. I am showing variations to the engine and it recognizes them. I am going some moves forward and backward and results are recognized by the engine.

It is possibility to save # (not persistent) so it is possible to return to the analysis later.

Rgds
Hood

[Uly]

@ernest: Rybka 3 had Persistent Hash, which is learning.

@hood:

>Rybka has preserve analysis option.

It's not learning. Because when you unload the engine, the engine has learned nothing.

>In the context of the current position analysis i assume that it is learning.

Then you have a twisted definition of learning. In your mind the Stockfish Gran series have learning, in mine, Learning has yet to be implemented in Stockfish.

>Rybka 3 has had preserve analysis as well . Does Rybka 4 have it ?

Yes, but:

1. Rybka 3's Preserve Analysis was not very effective because it slowed too much the engine when new positions appeared.

2. Rybka 4's Preserve Analysis doesn't have that drawback, so it is very powerful, however, as soon as you unload the engine, all the analysis preserved is deleted, so it's NOT learning.

[ernest]

@ernest: Rybka 3 had Persistent Hash, which is learning.

Learning on 1000 positions would mean saving 1000 Hash tables (maybe 1 TeraByte... )
Not very practical!...