Komodo 11.3 released

[Jesse Gersenson]

I think this release is more for the "real" users out there who use it for analysis etc than the ratings lists.

I enjoy playing against Komodo and usually play at depth=1 or 2 or 3. The new 'skill' uci setting is gives a much better game of chess. It's great fun.

[pilgrimdan]

Is default contempt of 16 adequate value against K11.2.2 in selfplay (assuming the latter use C=10) ?

Based on the estimated elo diff. and the formula in readme the optimum value for this test is Contempt = 1. But of course that wouldn't be fair if you use 10 in 11.2.2. A setting of Contempt = 13 would be the closest equivalent to 10 in the older version so it would be "fair". We raised it to 16 partly because SF now uses 20, so at least when we play Stockfish we are no longer handicapped, and against Houdini we're only very slightly handicapped at default levels.

can a search be done with 2 different comtempt values ... and then the lines be displayed ... seems like that would be an interesting engine that could do that ...

[peter]

Hi Larry!

Automatic backward analysis of games doesn't work at two different PCs of mine neither in Fritz14 nor in Shredder13- GUI, at both it's x64.exe, one running Win7Pro, the other one Win8.

Engine doesn't start computing after GUI- command is given.

[mjlef]

Is default contempt of 16 adequate value against K11.2.2 in selfplay (assuming the latter use C=10) ?

Based on the estimated elo diff. and the formula in readme the optimum value for this test is Contempt = 1. But of course that wouldn't be fair if you use 10 in 11.2.2. A setting of Contempt = 13 would be the closest equivalent to 10 in the older version so it would be "fair". We raised it to 16 partly because SF now uses 20, so at least when we play Stockfish we are no longer handicapped, and against Houdini we're only very slightly handicapped at default levels.

can a search be done with 2 different comtempt values ... and then the lines be displayed ... seems like that would be an interesting engine that could do that ...

Komodo has just one Contempt setting, which is used for that engine. Haveing a contempt value for white and one for black would basically just be the same as setting one to zero and the other to the difference. But in all the GUIs I have seen, you can create multiple engines with different configurations like Contempt and have them play each other or analyze with them. That would be useful to see what the opponent might do if it was trying to avoid draws, or seek them.

[mjlef]

I am not against a simple book learning system. I had that in my old program NOW. But it will just change its search based on what it found about a specific move in past games. So it could learn to avoid problems or remember a move in a specific position did well before. But nothing like learning what pawn configurations are best for specific pieces. So the leaning is only useful when you want to play the same lines over and over. If you are going to do that then you could also just save the hash table for reuse, which the program already does.

We would like to do something that learns generally and not just specific conditions. But if there is a demand for the simpler approach, we will listen.

[carldaman]

I am not against a simple book learning system. I had that in my old program NOW. But it will just change its search based on what it found about a specific move in past games. So it could learn to avoid problems or remember a move in a specific position did well before. But nothing like learning what pawn configurations are best for specific pieces. So the leaning is only useful when you want to play the same lines over and over. If you are going to do that then you could also just save the hash table for reuse, which the program already does.

We would like to do something that learns generally and not just specific conditions. But if there is a demand for the simpler approach, we will listen.

What the SF GTB PA learning file offers is a lot more than a book learning system. It is of great value to correspondence players and those users who do a lot of back-and-forth analysis of the openings.

For example, one could analyse his pet line of the Sicilian with Komodo. With learning enabled all the novelties and deep refutations found after long analysis sessions are stored in the learning file. When revisiting the same analysis some time later, all these key discoveries will be instantly remembered and propagated in the analysis, instead of having Komodo analyzing from scratch and remembering nothing as is the case now.

The same learning file will contain novelties and refutations in the Dutch, King's Indian, etc - any opening that the user likes to study. No discoveries will be lost and the wheel does not have to be reinvented again.

(This is probably all crystal clear to anybody who has used Stockfish GTB PA for analysis.)

[Vizvezdenec]

I don't think that SF contempt 20 is the same as komodo 20, also the latest dev version of sf uses 12+logarythmic dynamic contempt

Regarding SF "dynamic contempt", am I misunderstanding something? It seems to be saying that if you are winning, don't trade pieces, which is of course ridiculous. At least that's what it looked like a couple weeks ago.

Well, ridiculous or not, but it gains elo in selfplay and passed LTC SPRT easier than it did pass STC.
I think the main thing why it works can be formulated like this "no need to go to really advantageous endgame if you can wait and go for even more advantageous endgame".
Also it tries to trade away everything when position becomes bad, this is also maybe a place where elo gain comes from.

[carldaman]

I don't think that SF contempt 20 is the same as komodo 20, also the latest dev version of sf uses 12+logarythmic dynamic contempt

Regarding SF "dynamic contempt", am I misunderstanding something? It seems to be saying that if you are winning, don't trade pieces, which is of course ridiculous. At least that's what it looked like a couple weeks ago.

Well, ridiculous or not, but it gains elo in selfplay and passed LTC SPRT easier than it did pass STC.
I think the main thing why it works can be formulated like this "no need to go to really advantageous endgame if you can wait and go for even more advantageous endgame".
Also it tries to trade away everything when position becomes bad, this is also maybe a place where elo gain comes from.

It makes no sense to make so many trades in a bad position, unless those trades result in a drawn endgame, which is the exception rather than the rule.

[Vizvezdenec]

I don't think that SF contempt 20 is the same as komodo 20, also the latest dev version of sf uses 12+logarythmic dynamic contempt

Regarding SF "dynamic contempt", am I misunderstanding something? It seems to be saying that if you are winning, don't trade pieces, which is of course ridiculous. At least that's what it looked like a couple weeks ago.

Well, ridiculous or not, but it gains elo in selfplay and passed LTC SPRT easier than it did pass STC.
I think the main thing why it works can be formulated like this "no need to go to really advantageous endgame if you can wait and go for even more advantageous endgame".
Also it tries to trade away everything when position becomes bad, this is also maybe a place where elo gain comes from.

It makes no sense to make so many trades in a bad position, unless those trades result in a drawn endgame, which is the exception rather than the rule.

There are a lot of things that don't really make sense but that make engines play stronger...

[schack]

I'm seeing the same problem with auto-analysis in GUI15 from Chessbase.