Chessbase opening book

[pavel]

Under chessbase is it possible to make two engine play one opening line all the time (to see the influence of position learning). Let's say I was two enginesto play the E69 Fianchetto Kings indian, many gameswith engines alternating colors. Is it possible with chessbase or any other program?

I tried to create an opening book, before eng-eng match under chessbase but for some reason the engines are not using the moves from the book.

Thanks for any help.

[pavel]

Never Mind.

Got it.

Opening DB.

[Mike S.]

Got it. - Opening DB.

Exactly Seems you got it within only 9 minutes. Yes, if anyone wants to test an opening variation with learning engines, simply put it into a database and use it as an opening database without any books.

I did that some time ago, with Nimzo vs. Hiarcs and with Hiarcs vs. Shredder, always the same opening and 50 games each, and ALL games were different due to the position learning! Although, I must say that the results didn't provide the perfect conclusions for me either...

[pavel]

Nimzo was one sweat engine. But Nimza-Hiarcs-Shredder all has/had positional learning if I am not wrong. I think the ideal way to test the influence of positional learning is to play an opponent engine without positional learning.

Position learning has always interested me in chess engine, which is why I loved YACE. Did all the leading engine authors gave up on position learning?

I am currently testing Spike with Rybka-mp on E69, Rybka is tearing apart Spike, but I will let it run 200-300 games and see what happens.

cheers.

[Ovyron]

Did all the leading engine authors gave up on position learning?

Rybka 3 will have Persistent Hash Learning.

[pavel]

I will buy a new version of Rybka only if it can read minds...

j/k ;)

I wonder what Persistent Hash Learning actually means?

[Mike S.]

I think the ideal way to test the influence of positional learning is to play an opponent engine without positional learning.

Maybe, but that was not the purpose of the tests I mentioned. I wanted to test the quality of specific opening continuations, and (hopefully) find the best continuations, from several considerable continuations. But as I remember it, due to the learning functions the engines changed their mind so often that no clear conclusion was possible. Sometimes, Shredder would even deviate from a game it had won and then only draw or lose after the alternative.

I guess, next time when I do opening research like that, I will rather stick to "conservative" analysis or to "deep analysis" functions, instead.

[Ovyron]

I wonder what Persistent Hash Learning actually means?

It means that the engine keeps her hash from all its games. So, analyzing a position for 1 hour, restarting the computer and analyzing for another hour is the same than analyzing it for 2 hours.

The learning is cumulative and it has a high potential. For example, you can let Rybka play a very long game from a position, and it's expected that it would play much better the beginning of that position in bullet time controls (as it will be using the information from the long game.)

The main difference from other implementations is that some others are based on game results. That means that bad moves that won a game are learned as good, and good moves that lost a game are learned as bad. Persistent Hash doesn't have this problem.

Also, it's expected that hash files will be able to be merged, so I can upload my learn file somewhere, and you could download it and merge it with your learn file at home for stronger learning.

[pavel]

This sounds very interesting, actually this is close to something I asked dieter author of YACE to implement, when he asked open question to everyone what features he should implement in YACE, in winboard forum.

I hope it gets implemented in the upcoming version. That would be nice.