Komodo 11.3 released

[CMCanavessi]

11.3 would be equivalent to one of these test version: 2023? 2016?

http://www.talkchess.com/forum/viewtopi ... 5&start=14

[carldaman]

Hi Larry,

With the Elo gains slowing down, I can only think of one surefire antidote: learning!

Think about it, a Komodo that learns will continue to add Elo from its learning. It is a back door to gaining Elo and can be advertised as such -- Komodo, the top engine that learns and keeps getting stronger the more YOU use it!

Anyone who doesn't have such a learning-equipped Komodo will be dying to have it, and sales would skyrocket. With learning enabled it should surpass SF and Houdini before long.

Yes, the rating lists may not test such a version, but the undeniable fact will be that Komodo will be empowered to inevitably become the strongest in terms of both playing and analysis just by routine use on the users' PC (!)

Don't forget, there are some who are still using a 4-5 year old Stockfish version with permanent hash learning, that used to be developed by Jeremy Bernstein, all because of how powerful its learning is! Imagine if you could offer a brand new Komodo with such a feature...

Once implemented, this feature will remove much of the pressure to gain Elo by normal painstaking means. It will be far less stressful for you and Mark and a tremendous boost to your customers. Remember, no one else if offering this feature (in a top engine).

Regards,
CL

[lkaufman]

Hi Larry,

With the Elo gains slowing down, I can only think of one surefire antidote: learning!

Think about it, a Komodo that learns will continue to add Elo from its learning. It is a back door to gaining Elo and can be advertised as such -- Komodo, the top engine that learns and keeps getting stronger the more YOU use it!

Anyone who doesn't have such a learning-equipped Komodo will be dying to have it, and sales would skyrocket. With learning enabled it should surpass SF and Houdini before long.

Yes, the rating lists may not test such a version, but the undeniable fact will be that Komodo will be empowered to inevitably become the strongest in terms of both playing and analysis just by routine use on the users' PC (!)

Don't forget, there are some who are still using a 4-5 year old Stockfish version with permanent hash learning, that used to be developed by Jeremy Bernstein, all because of how powerful its learning is! Imagine if you could offer a brand new Komodo with such a feature...

Once implemented, this feature will remove much of the pressure to gain Elo by normal painstaking means. It will be far less stressful for you and Mark and a tremendous boost to your customers. Remember, no one else if offering this feature (in a top engine).

Regards,
CL[/quote

I am not opposed to the idea, but simply learning what might be a better move in specific positions isn't much different from just using a very good opening book which is regularly updated. Now learning in the alpha-zero sense is another matter entirely, but we don't yet know how practical that might prove to be.

[Ovyron]

Think about it, a Komodo that learns will continue to add Elo from its learning. It is a back door to gaining Elo and can be advertised as such -- Komodo, the top engine that learns and keeps getting stronger the more YOU use it!

They could go beyond that...

- Provide a tool to merge learn files
- Provide a way for users to send learn files
- Have a big learn file on the komodo server with everything all users have sent
- Allow users to download the learn file and merge it with the one they have automatically

Now, all Komodos everywhere are able to benefit from what other users have made them learn.

Apparently, the biggest obstacle about this is the users themselves (they'd download the learning but wouldn't provide learning of their own, as they want to keep it off the hands of their opponents), but Chessbase Cloud had limited success, so something like this could work.

[Ovyron]

I am not opposed to the idea, but simply learning what might be a better move in specific positions isn't much different from just using a very good opening book which is regularly updated.

We're not talking about learning for games, we're talking about something like Rybka 3 Persistent Hash, or Critter 1.6a's session file.

Learning is extremely useful for analysis, and great that you can unload the engine and when you load it again you start from where you were, and Komodo remembers everything analyzed which speeds up the analysis process and may allow Komodo to find transpositions to stuff already analyzed.

Currently users are having to store their analysis on the GUI, on trees, on books, or on graphs, because the engine forgets everything analyzed, but engines with learning like Learning Stockfish makes such things something of the past, because I don't need to write down by hand on my file all the analysis provided, and don't need to rely on something like "Aquarium's IDEA" to store the engine's analysis.

And in a very few space, I used Stockfish PA GTB that has learning for 4 years (ceased a month ago as it was superseded) and the invaluable learning file only grew up to 8.5MB.

I wonder if I'd need to make a demonstration of how Learning makes a difference and allows the engine to find the best moves on chess posisions faster. I kept using Rybka 3 years after there were engines hundreds of elo stronger just because its learning mechanism allowed it to provide much better analysis. Say, my opponent played an unexpected move that I didn't have covered in my analysis file, and all engines would scratch their heads and start from scratch, while Rybka 3's Persistent Hash already had stored the best move and it was easy to convince other engines of this, saving lots of time having to find that move. Komodo can't do that, as it forgets everything analyzed.

[carldaman]

Think about it, a Komodo that learns will continue to add Elo from its learning. It is a back door to gaining Elo and can be advertised as such -- Komodo, the top engine that learns and keeps getting stronger the more YOU use it!

They could go beyond that...

- Provide a tool to merge learn files
- Provide a way for users to send learn files
- Have a big learn file on the komodo server with everything all users have sent
- Allow users to download the learn file and merge it with the one they have automatically

Now, all Komodos everywhere are able to benefit from what other users have made them learn.

Apparently, the biggest obstacle about this is the users themselves (they'd download the learning but wouldn't provide learning of their own, as they want to keep it off the hands of their opponents), but Chessbase Cloud had limited success, so something like this could work.

Yes, the possibilities are almost endless that way, but at the very least having something like SF GTB PA, [which used to be maintained by Jeremy Bernstein, a friend of Team Komodo (!)] would be more than half the battle.

[carldaman]

Hi Larry,

With the Elo gains slowing down, I can only think of one surefire antidote: learning!

Think about it, a Komodo that learns will continue to add Elo from its learning. It is a back door to gaining Elo and can be advertised as such -- Komodo, the top engine that learns and keeps getting stronger the more YOU use it!

Anyone who doesn't have such a learning-equipped Komodo will be dying to have it, and sales would skyrocket. With learning enabled it should surpass SF and Houdini before long.

Yes, the rating lists may not test such a version, but the undeniable fact will be that Komodo will be empowered to inevitably become the strongest in terms of both playing and analysis just by routine use on the users' PC (!)

Don't forget, there are some who are still using a 4-5 year old Stockfish version with permanent hash learning, that used to be developed by Jeremy Bernstein, all because of how powerful its learning is! Imagine if you could offer a brand new Komodo with such a feature...

Once implemented, this feature will remove much of the pressure to gain Elo by normal painstaking means. It will be far less stressful for you and Mark and a tremendous boost to your customers. Remember, no one else if offering this feature (in a top engine).

Regards,
CL[/quote

I am not opposed to the idea, but simply learning what might be a better move in specific positions isn't much different from just using a very good opening book which is regularly updated. Now learning in the alpha-zero sense is another matter entirely, but we don't yet know how practical that might prove to be.

Thanks for your reply, Larry. A version of Komodo that does not forget its analysis discoveries, similar to SF GTB PA by Jeremy Bernstein, should not be underestimated.

Another advantage of this method is that the learning file is small and tends to grows slowly, only adding stuff it discovers past a certain depth, which can be adjusted by the user. The same analysis file can thus be used in any position, for all analysis purposes. Only the stuff discovered at greater depths will get stored and remembered. This also helps with backwards propagation of scores in back-and-forth analysis.

Now who said anything about alpha-zero?
A precedent exists, and was successfully implemented in an old Stockfish fork (see above). No alpha-zero type of learning is required.

[Ovyron]

Yes, the possibilities are almost endless that way, but at the very least having something like SF GTB PA, [which used to be maintained by Jeremy Bernstein, a friend of Team Komodo (!)] would be more than half the battle.

Yeah, though I'll just drop it here: everyone would alreay have something like that if it wasn't for Stockfish's licence, nobody is able to take GTB PA's code and plug it into their engine and sell it, if someone implements it in their closed source engine, it's against the terms of use to distribute it. If Komodo implemented GTB PA's learning into Komodo they'd need to open source the engine! It's all crazy stuff! People are being forced to reinvent the wheel each time!

Just saying...

[JJJ]

Congrats for your release Larry and Mark.
Did you try to play against Komodo at Magnus Carslen level Larry ? Or did someone you know try ?

[carldaman]

Yes, the possibilities are almost endless that way, but at the very least having something like SF GTB PA, [which used to be maintained by Jeremy Bernstein, a friend of Team Komodo (!)] would be more than half the battle.

Yeah, though I'll just drop it here: everyone would alreay have something like that if it wasn't for Stockfish's licence, nobody is able to take GTB PA's code and plug it into their engine and sell it, if someone implements it in their closed source engine, it's against the terms of use to distribute it. If Komodo implemented GTB PA's learning into Komodo they'd need to open source the engine! It's all crazy stuff! People are being forced to reinvent the wheel each time!

Just saying...

They could write their own code based on the same idea. They would only have to open source any code taken directly from SF GTB PA, and I figure they won't take any, so that's a moot point. Unless I'm missing something.

Edit: No one has done something similar, because it's too much work for a non-commercial endeavor. Houdini's feature was buggy and was dropped unfortunately, and Komodo's authors have said before that they'd add learning if there was sufficient interest.