Any news of a Komodo update in sight?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Tobber
Posts: 379
Joined: Fri Sep 28, 2012 3:53 pm
Location: Sweden

Re: Any news of a Komodo update in sight?

Post by Tobber » Mon Dec 11, 2017 5:19 pm

Werewolf wrote:
shrapnel wrote:
Werewolf wrote:Alpha Zero won on much more powerful hardware.
Ahh..Werewolf, we meet again.
So you've convinced yourself that AlphaZero won only because of superior hardware !
Awesome !
Did you even bother going through the Games ?
You're living proof that one can convince oneself of ANYTHING, provided one believes strongly enough, despite all evidence to the contrary !
You believe, what you want to believe.
I wonder if you've looked at the evidence.

Alpha Zero is impressive, did I say otherwise?
But Stockfish was CRIPPLED with no opening book, 1 GB of RAM (even my laptop has more!) and an awful time control not to mention they used old Stockfish 8!

Add up all these disadvantages AND factor in the superior hardware of AlphaZero and it didn't win by that much.

Give me the same amount of money they spent on their hardware and I bet I could equal it with SF (with some help from the SF team).
Please read the document again, they have not said anything about RAM, they used 1GB of hash size and 64 threads. Arguments from ignorance are not worth much, you seem not equipped to discuss this topic.

/John

Werewolf
Posts: 1349
Joined: Thu Sep 18, 2008 8:24 pm

Re: Any news of a Komodo update in sight?

Post by Werewolf » Mon Dec 11, 2017 5:42 pm

Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?

Leo
Posts: 996
Joined: Fri Sep 16, 2016 4:55 pm
Location: USA/Minnesota
Full name: Leo Anger

Re: Any news of a Komodo update in sight?

Post by Leo » Mon Dec 11, 2017 5:58 pm

Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?
To limit the strength of SF.
Advanced Micro Devices fan.

Tobber
Posts: 379
Joined: Fri Sep 28, 2012 3:53 pm
Location: Sweden

Re: Any news of a Komodo update in sight?

Post by Tobber » Mon Dec 11, 2017 6:12 pm

Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?[/quote

Since you compared it to your laptop memory it was more likely a "thinko" than a typo. How much it crippled SF I don't know but I guess someone have tested it somewhere. According to Komodo 128 MB is enough for rapid games, how much is needed for 1 minut/move I don't know but I'm not convinced it would have a major impact on the strength. The hash size is related to move time.

/John

abulmo2
Posts: 311
Joined: Fri Dec 16, 2016 10:04 am
Contact:

Re: Any news of a Komodo update in sight?

Post by abulmo2 » Mon Dec 11, 2017 6:13 pm

Leo wrote:
Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?
To limit the strength of SF.
This have to be proved. I would like to see a SF 8 vs SF 8 (or Komodo/Houdini) match with different hash sizes to see the impact of the hash size to the elo, on the machine used by the alphaZero team. Unfortunately, I do not have a computer powerful enough to do any test, so I cannot do anything meaningful myself.
Richard Delorme

User avatar
Laskos
Posts: 10949
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: Any news of a Komodo update in sight?

Post by Laskos » Mon Dec 11, 2017 6:43 pm

abulmo2 wrote:
Leo wrote:
Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?
To limit the strength of SF.
This have to be proved. I would like to see a SF 8 vs SF 8 (or Komodo/Houdini) match with different hash sizes to see the impact of the hash size to the elo, on the machine used by the alphaZero team. Unfortunately, I do not have a computer powerful enough to do any test, so I cannot do anything meaningful myself.
See here, for example:
http://www.talkchess.com/forum/viewtopi ... t&start=20

In the conditions of the match with A0, probably the necessary Hash was 32 or 64GB, and by looking at the link, it seems, SF8 had some 10% time-to-depth disadvantage. How it translates in Elo points there I don't know, Elo becomes a bit obsolete when the score is 28-0 with 72 draws. Maybe with 64GB Hash SF8 would have gotten 1-2 more draws. This Hash issue is a bit irrelevant, when we are talking of such hard to compare hardware used by A0 and SF8.

Werewolf
Posts: 1349
Joined: Thu Sep 18, 2008 8:24 pm

Re: Any news of a Komodo update in sight?

Post by Werewolf » Mon Dec 11, 2017 9:39 pm

Tobber wrote:
Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?[/quote

Since you compared it to your laptop memory it was more likely a "thinko" than a typo. How much it crippled SF I don't know but I guess someone have tested it somewhere. According to Komodo 128 MB is enough for rapid games, how much is needed for 1 minut/move I don't know but I'm not convinced it would have a major impact on the strength. The hash size is related to move time.

/John
It wasn't a "thinko" since I had already posted about the hash size a few days earlier here:

http://www.talkchess.com/forum/viewtopic.php?t=65935

Rather, it was on my laptop "I typically use 2 GB hash" (i.e. more than they used).

Given they had 64 threads the size of the hash is woeful (on my 24 core 1 GB fills very quickly!) and there's no excuse for using so little, except it keeps nps high while limiting strength.

Am I saying changing the hash would have saved the match? NO
But it would have helped.

The T/C, opening book, and version of engine use all contribute as well.
Last edited by Werewolf on Mon Dec 11, 2017 9:41 pm, edited 2 times in total.

Werewolf
Posts: 1349
Joined: Thu Sep 18, 2008 8:24 pm

Re: Any news of a Komodo update in sight?

Post by Werewolf » Mon Dec 11, 2017 9:39 pm

Leo wrote:
Werewolf wrote:Sorry that was a typo - I meant 1 GB of Hash.

And why did they use such a tiny amount?
To limit the strength of SF.
Precisely.

shrapnel
Posts: 1340
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: Any news of a Komodo update in sight?

Post by shrapnel » Tue Dec 12, 2017 12:26 pm

Laskos wrote:How it translates in Elo points there I don't know, Elo becomes a bit obsolete when the score is 28-0 with 72 draws. Maybe with 64GB Hash SF8 would have gotten 1-2 more draws. This Hash issue is a bit irrelevant, when we are talking of such hard to compare hardware used by A0 and SF8.
This is the part which seems to be evading Werewolf's attention.
He is just simply desperately looking for excuses to explain the Rout of Stockfish. He will accept everything except the relevant FACT that there has been a quantum jump in computer chess, rendering Alpha Beta type Chess Engines obsolete.
To be fair, its been too sudden a shock for him and others like him and he will take time to adjust to the fact that there is a new Alpha in Town !
Of course, DeepMind not releasing more information and games since then for what ever reason, will only encourage his type.
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

Michael Sherwin
Posts: 3196
Joined: Fri May 26, 2006 1:00 am
Location: WY, USA
Full name: Michael Sherwin

Re: Any news of a Komodo update in sight?

Post by Michael Sherwin » Tue Dec 12, 2017 12:50 pm

mjlef wrote:
shrapnel wrote:
Jesse Gersenson wrote:If people want learning in Komodo, let your voice be heard; Mark and Larry are open to feature requests, especially those requested by a lot of people.
'.
So Reinforcement Learning will be introduced by Komodo team (if they are capable of it) only on REQUEST.
Looks like the thrashing stockfish received still hasn't convinced them.
*** Unbelievable.
Of course we listen to requests. But I do not think the Romi learning is anything like the learning Google/DeepMind did. They used 5000 special TensorFlow Processing units (TPUs), each costing thousands of dollars. Right now, this is way beyond our resources. Romi learning is most likely simply saving important positions in a persistant hash. In future games, these are reloaded into the main hash table, so the new game, if ir encouters one of these positions, alread has deep search information ofr them. The thing is, this helps not at all if a different lien is played. My old program NOW had this feature, but I did it not so much to make it stronger, but instead to avoid losing lines during tournaments. You can think of it as a self-correcting book, which benefits if the same line it tried on Komodo.

Larry and I often discuss Monte Carlo Tree Search, and are interested in trying this. We have also discussed uses for neural networks. Small nns could be useful in present PCs, but the massive nn used in AlphaGo Zero is currently beyond what we, and most chess engine users can afford.

We listen, and try to add what we think people want. But we do not have endless resources. We can afford to buy roughly 1 new server each year. It is not a matter of being "convinced". We just cannot afford it currently. Hopefully, GPUs in graphics cards will get faster and perhaps nn hardware will get added to future PCs or at least be much less expensive.
Mark Lefler was kind enough to ask me in a pm to error check the above! :D
But I do not think the Romi learning is anything like the learning Google/DeepMind did.
There is circumstantial evidence that it is similar in nature. Reinforcement learning involves accumulated rewards (and penalties) for every position in a database. It could be a persistent hash like suggested.
Romi learning is most likely simply saving important positions in a persistant hash.
Or the way Romi actually does it in a tree of all played games. This is superior to a persistent hash as only the subtree from the current position is loaded into the game hash. This has the advantage that only useful information ends up in the game hash.
The thing is, this helps not at all if a different lien is played.
Not true. Rewards and penalties are greater near the leaves and over time are back propagated to the root. Since every node is a root to a subtree, every move benefits from backpropagation. This results in a meaningful differentiation resulting in a gradual determination of which move gives better results. So for example at the actual root Romi will settle on say 1.e4 or 1.d4 etc. as being best and will always play that opening move. Just like AlphaZ always played 1.d4.
You can think of it as a self-correcting book, which benefits if the same line it tried on Komodo.
Philosophically I would argue that what RomiChess does has nothing to do with a book. In a book an engine usually can choose randomly between acceptable moves. In RomiChess only the learned best move is played in its Monkey See Monkey Do "book". That is separate from Romi's reinforcement learning. It just so happens that the stored tree of all Romi's games can handle both a "book" and the rewards/penalties for reinforcement learning.

Thanks Mark for inviting me to share some details! :D
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through

Post Reply