Page 4 of 4

Re: Incredible Stockfish Patch

Posted: Fri Dec 29, 2017 1:00 am
by Leo
Many are looking forward to this. I always liked learning programs and this will be the best one.

Re: Incredible Stockfish Patch

Posted: Fri Dec 29, 2017 2:35 am
by pilgrimdan
Thomas Zipproth wrote:
Leo wrote:Once Brainfish and the Cerebellum Library (not yet released) is available for purchase then we can make our own learning library.
Thanks for your support. You are right, Cerebellum and the Sirius Gui will be published.
The delay is due to a lack of time , but currently I'm working on the project. I will publish a new release date as soon as I can predict it reliable.

Btw, I have updated the Cerebellum AlphaZero Analysis Page to make it more readable.
http://www.zipproth.de/Brainfish/Cerebe ... aZero.html

Thomas
thank you very much for updating the Cerebellum a0 Analysis Page ...

Re: Incredible Stockfish Patch

Posted: Fri Dec 29, 2017 2:47 am
by leavenfish
Ovyron wrote:
carldaman wrote:Stockfish and Houdini would be worth much less in comparison, if that were to happen. :wink:
Indeed, Houdini attempted a learning feature, but it was so bad nobody used it and it was removed from later versions....etc
Why don't other top free engine authors implement learning? They're too busy chasing elo, what with them not even fully implementing UCI? etc
I am not going to disagree with you there. I've said it...I'll say it again, most of these 'engine programmers' are spending their time being little more than guys who tweak here, tweak there, test and hope to hit upon something. Just my humble opinion...people can throw their tomatoes if they like...

Really, these engines are already so good, they should try to do something with them. But it's easier to tweak away...

Re: Incredible Stockfish Patch

Posted: Fri Dec 29, 2017 3:56 am
by Ovyron
leavenfish wrote:I am not going to disagree with you there. I've said it...I'll say it again, most of these 'engine programmers' are spending their time being little more than guys who tweak here, tweak there, test and hope to hit upon something. Just my humble opinion...people can throw their tomatoes if they like...
Well, I don't know how controversial what I'm going to say is going to sound, but particularly for Houdini, here's my report of the engine's analysis quality for correspondence games (and I claim this would hold for any kind of analysis.)

Houdini 1.0 - Usable for its time, due to the weakness of other engines, but weaker than later Houdinis.
Houdini 1.5 - By its time, better analysis engines existed.
Houdini 2.0 - Garbage
Houdini 3.0 - Garbage
Houdini 4.0 - Garbage

Houdini 6.0 - Fantastic!

That is, up until Houdini 4.0, the engine was providing analysis that was inferior to that of Rybka 4.1, Rybka is sitting at place 18th on the rating list... Houdini was providing garbage analysis worth of the 19th top engine of the lists.

Fritz 15 was released, and Rybka 4 was obsoleted, because Fritz is just Rybka with a few changes (now Rybka gets to appear twice on the rating lists :D )

Now, I don't know what Robert Houdart did for Houdini 6, but it wasn't just a "tweak, tweak" to gain elo. This Houdini not only provided better analysis than Rybka...

I have found, on a regular basis, that Houdini 6 provided better analysis than Komodo (10), and better evaluation than Stockfish dev.

Komodo keeps suggesting super-solid variations with long term planning that get destroyed by Houdini attacks. I'm winning two games thanks to Houdini.

What do I mean with evaluation? Something like this:

You analyze the positions interactively, and get three furure positions that Stockfish reaches with the following evaluations:

Position A: Score 0.16
Position B: Score 0.18
Position C: Score 0.19

You may think the move that leads to Position C is slightly better.

What does Houdini 6 say?

Position A: Score 0.08
Position B: Score 0.42
Position C: Score 0.06

Holy macaronni! 0.42? What is going on? You show Houdini variations to Stockfish, and Stockfish variations to Houdini. Stockfish was missing some defensive lines, so A and C take a hit...

Position A: Score 0.14
Position B: Score 0.18
Position C: Score 0.13

Stockfish isn't impressed by Houdini's continuation for B, and actually refutes it, but it's still clear favorite for Houdini...

Position A: Score 0.08
Position B: Score 0.24
Position C: Score 0.06

You make the move. After the opponent replies you find that A and C are still reachable with tanspositions, but it's your last chance to get to them, so you analyse deeper, what do engines say?

Stockfish
Position A: Score 0.11
Position B: Score 0.20
Position C: Score 0.09

Houdini
Position A: Score 0.00
Position B: Score 0.30
Position C: Score 0.04

Position B still clear favorite, the others going to hell, while initially Stockfish was thinking they were more or less equal, but B was Houdini's favorite.

It doesn't matter that Houdini was overoptimistic with B, what matters is it knew it was much better than the alternatives, and knew the others were drawish, that's better eval.

Of course you still need Stockfish, as Houdini wasn't seeing some better moves the opponent could play, but I have found Houdini 6 analysis indispensable for analysis, because you never know if Stockfish's second or third best is actually clearly the best variations.

So, at least in this case I'm glad the author spent its time improving... whatever was improved to jump from garbage analysis to fantastic analysis, instead of fixing Houdini's learning feature. As long as it's not just elo chasing...