Komodo 14 released.

[lkaufman]

@ Larry

Just purchased K14....
I was wondering if the MC version running around 4 kn/s is correct or is it because I am using my CPU (AMD Threadripper 1950x) instead of my GPU (2080Ti) ?

Thanks for the engines!

rgds

Komodo (at least publicly available Komodo) does not yet use GPU, so yes, this is correct. Each MCTS node is based on hundreds or thousands of regular nodes, so that 4 ks/s really equates to millions of normal nodes per second.

[Lion]

Thanks for your response!

Reading "in between the lines", should we understand there may be at some point a GPU MC version? Should we exect an increase of speed with such a configuration/setup compared to CPU MC?

rgds

[lkaufman]

Thanks for your response!

Reading "in between the lines", should we understand there may be at some point a GPU MC version? Should we exect an increase of speed with such a configuration/setup compared to CPU MC?

rgds

It's likely that we will have an MC version that uses NN in the next year, and if so it will probably play much faster/stronger with GPU than without. But we're pursuing multiple leads so I don't know what will end up working well enough for release, we may be surprised ourselves.

[Nordlandia]

Is it sensible to launch selfplay with armageddon enabled on both sides ?

Let's take an extreme case, for instance in the drawkiller starting position:

Maybe contempt values need to be adjusted to offset the first move advantage

[d]knbrqbnr/ppp1pppp/3p4/8/8/4P3/PPPP1PPP/RNBQRBNK w - - 18 11

[lkaufman]

Is it sensible to launch selfplay with armageddon enabled on both sides ?

Let's take an extreme case, for instance in the drawkiller starting position:

Maybe contempt values need to be adjusted to offset the first move advantage

[d]knbrqbnr/ppp1pppp/3p4/8/8/4P3/PPPP1PPP/RNBQRBNK w - - 18 11

In theory it should work properly, but I haven't tested it to be sure. There should be no need to adjust contempt for the first move advantage, at least it shouldn't need a significant adjustment. Now if someone starts two pawns up or so then this might be an issue.

[Krzysztof Grzelak]

I'm sorry for asking. I have a reading for Komodo developers when there is a new version of Komodo.

[lkaufman]

I'm sorry for asking. I have a reading for Komodo developers when there is a new version of Komodo.

We have some improvements since Komodo 14, but not yet enough to justify Komodo 14.1. We can never really anticipate when we'll discover something good, but I would bet that we'll have enough some time in July. Things are happening with Komodo NN too, but I can't predict how long that will take to produce something we can release.

[Andrew]

Something for Larry, or any of the LC0 people to comment on. I noticed recently on this page

https://www.chessvariants.com/programs.dir/zillions/

Mark Lefler is mentioned, is this the Komodo Mark?

What's interesting is that they developed a system of rule files for all sorts of chess variants
(and beyond).

I'd be really keen to see some sort of NN learning program developed where you can load up a
rule file such as one of these, enter a few parameters (such as network size etc) and click go!

Especially for some of the smaller board variants eg
https://en.wikipedia.org/wiki/Los_Alamos_chess and several here:
https://en.wikipedia.org/wiki/Minichess

I'd hope with these smaller variants learning games would run much more quickly, and strength would
improve more quickly. Which would make it great for people to tinker around with on more
modest computers.

Andrew

[mjlef]

Something for Larry, or any of the LC0 people to comment on. I noticed recently on this page

https://www.chessvariants.com/programs.dir/zillions/

Mark Lefler is mentioned, is this the Komodo Mark?

What's interesting is that they developed a system of rule files for all sorts of chess variants
(and beyond).

I'd be really keen to see some sort of NN learning program developed where you can load up a
rule file such as one of these, enter a few parameters (such as network size etc) and click go!

Especially for some of the smaller board variants eg
https://en.wikipedia.org/wiki/Los_Alamos_chess and several here:
https://en.wikipedia.org/wiki/Minichess

I'd hope with these smaller variants learning games would run much more quickly, and strength would
improve more quickly. Which would make it great for people to tinker around with on more
modest computers.

Andrew

Andrew,

I am coauthor of Zillions of games. Those sites post .zrf (Zillions Rules Files) which are used by Zillions of Games. Zillions loads these files which describe board topology, pieces movements, the goal of the games and GUI information like what bitmaps to use for pieces, then derives a search and evaluation function for that game. It is a universal board game playing program, with thousands of games and variants available from the chess variant pages and zillions-of-games.com.

I think designing a NN to be used with Zillions is a great idea. Zillions does support external .DLL file which it can use in place of its search and evaluation. A sample is here: http://www.zillions-of-games.com/progsample.html. I am busy full time with Komodo, but maybe someone with the skills needed to develop such a NN systm can work on it.

This might interest you: Since many games like checkers suffer from zugzwang, I could not use nullmove in Zillions. Instead I came up with a scheme that used move number and evaluation to reduce later moves. We call this Late Move Reductions now. And his was back in 1997.

Mark

[Ovyron]

Honestly, what we need is an open source version of Zillions, or at least a free one. Or one that does what it does but it's free.

Then we implement NNs on that one, or it's going to be a lost cause.