Dragon by Komodo 3.3 "MCTS"

[Frank Quisinsky]

Wow, I learn every day new things!
So strong, Dragon by Komodo 3.3 MCTS.
I had never noticed that, at the moment a test-run is still running.
I am wonder ...

Congratulation for ths strength with MCTS.

[Modern Times]

But much weaker than standard mode, maybe -50 Elo.

[Frank Quisinsky]

In the evening I know more, but I believe not more as 50 Elo weaker.

Komodo 14.1 HCE / Komodo 14.1 HCE MCTS will be follow.
Here is interesting, also 50 Elo differents or more.

Thats really interesting for me, because for me totaly new knowledge.
Apparently, I’ve often dozed off while testing engines in the past.

Ah, you gave the information with 50 Elo in your message.
I didn't notice that at first.

[Tibono]

Hi,
I let both versions play in tournaments 40/10 (single core, ponder on, 15 secs avg.); so far 140 games achieved by the standard NNUE and 100 games by the MCTS version; against various opponents. Currently in my private rating list, the Elo score is -37 for the MCTS. Also, 3.3 MCTS is stronger by a dozen Elo points compared to 2020' Dragon (90 played games).

[Modern Times]

Look at CCRL blitz for a comparison.

[Frank Quisinsky]

Hi Eric,

is a small challenge between to test also the older HCE versions again. Many of the engines are with newer compiles stronger as we expect.
I think HCE engines will become much more popular than they are today, because many of them far exceed the maximum human level. In that respect, it remains fascinating. Could be a little bit to compare with the older chess computers. Today the older chess computer have a lot of fans. NN is a very own challenge for me, interesting of course, but thats not all.

So all is interesting also your results.
All is interesting, to each time ...

Furthermore, MCTS is interesting!
Monty for an example or Jackal.
But I never expect such a strong level from Dragon by Komodo 3.3 MCTS.

Best
Frank

[Frank Quisinsky]

Hi Modern-Times,

the problem with rating systems, like CCRL or CEGT are ...
Different hardware, different opponents and information to Elo only.
Bad for stats because after all I know both systems are working with adjudication = on.
Private engines are available in these lists, engines the name of the programmers are unknown and so on, unnecessary results, such as engine versions with more than one core. A lot more ... will not go in details.

Its much more interesting to find out information in self work as to look on such systems.
That's my personal opinion.

But after all, I like CCRL and of course CEGT for other reasons because without the list I have clearly lesser information to collect information ... for an example ... all the engines in my "newer" link selection. So, for different reasons I like also to look in CCRL or CEGT, most I like the small tournaments from Graham. All the years for me the "TalkChess-Event" number 1.



But if I like to find out a bit ...
With my own conditions and in live mode on my own PCs.
Thats more interesting, I do that since more as 30 years, a long time before CEGT and CCRL started, you know. I am full of ideas and own build expertise to do this and that.

A good example can be found on Elostat readme ... an old Winboard list.
I created such things in chess computer times, and looking here more as on SSDF, because I created my own stats about it, to every time if I like to do it in my life. Thats so easy to do that.

Best
Frank

And exactly this one should be the message for all the others:
Create your own test and have fun with it !!!
The journey is the destination, you learn much more as to looking in a ready list.

[Eelco de Groot]

This is Komodo in CCRL Fischer Random Chess as an example of the independent testing by 'Modern Times' but he wants to remain a bit anonymous I suppose. The difference is 95 points with Komodo 3.3 without using Monte Carlo. Just trying to help Frank a bit if he does not know ... Frank has been here much longer than I so I think he does know very well...

[Modern Times]

This is Komodo in CCRL Fischer Random Chess as an example of the independent testing by 'Modern Times' but he wants to remain a bit anonymous I suppose. The difference is 95 points with Komodo 3.3 without using Monte Carlo. Just trying to help Frank a bit if he does not know ... Frank has been here much longer than I so I think he does know very well...

No that is incorrect, I was referring to CCRL blitz standard chess.

https://computerchess.org.uk/404/cgi/co ... librate=no

But not many games.

I ceased FRC a few years ago and my association with CCRL ended a few years ago as well. They do their own thing, I do mine.

[Frank Quisinsky]

In my test it is 52 Elo ... Dragon by Komodo 3.3 to Dragon by Komodo 3.3 MCTS.
Really very very strong the MCTS version.