Is Carlsen's approach similar to that of AlphaZero?

Traditional chess games and chess topics in general

Moderators: Dan Honeycutt, Sam Hull, fern

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
serpico
Posts: 1
Joined: Fri Dec 14, 2018 1:12 pm
Full name: Riccardo Musso

Is Carlsen's approach similar to that of AlphaZero?

Post by serpico » Tue Apr 30, 2019 5:59 pm

In the last few months I've read on some blogs that Carlsen's approach in this 2019 is similar to that of AlphaZero. Could there be any truth in this?

jp
Posts: 645
Joined: Mon Apr 23, 2018 5:54 am

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by jp » Wed May 01, 2019 9:27 am

No, it isn't. Carlsen has been developing his all-round understanding of the game for some time now. It's showing huge results now, which it didn't in the WC match.

Comparing Carlsen & A0 is insulting to Carlsen and to human GMs (although it's some of those same human GMs who speculate about similarity). A0 is just using brute calculational force. This simple fact is somehow swept aside by all the propaganda. (We also do not have access to all of A0's games.)

Carlsen has deep understanding. Engines of any kind have brute calculating power.

koedem
Posts: 61
Joined: Fri Mar 18, 2016 9:45 pm

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by koedem » Wed May 01, 2019 12:46 pm

jp wrote:
Wed May 01, 2019 9:27 am
A0 is just using brute calculational force. This simple fact is somehow swept aside by all the propaganda. (We also do not have access to all of A0's games.)

Carlsen has deep understanding. Engines of any kind have brute calculating power.
While that is true of course, it uses that power for a very much higher level of play. As for the tests with Leela that I have seen, in blitz even with only one node per move (which literally is about as little as you can calculate) it still is very much competitive with top GMs.

jp
Posts: 645
Joined: Mon Apr 23, 2018 5:54 am

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by jp » Wed May 01, 2019 1:42 pm

koedem wrote:
Wed May 01, 2019 12:46 pm
While that is true of course, it uses that power for a very much higher level of play. As for the tests with Leela that I have seen, in blitz even with only one node per move (which literally is about as little as you can calculate) it still is very much competitive with top GMs.
Have you looked at the thread on this topic in the main forum? (I can put a link here if you haven't.) That is a dubious claim (that it is competitive at N=0 with top GMs or even medium GMs). Maybe at bullet it is, because there the poor humans barely have time just to move the pieces without thought.

I understand what you mean by "literally about as little as you calculate", but of course the real "as little as you can calculate" is N=0 with a NN size that is as small as possible, not a 2MB one.

jp
Posts: 645
Joined: Mon Apr 23, 2018 5:54 am

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by jp » Wed May 01, 2019 2:42 pm

Above, it should say 52MB, not 2MB.

koedem
Posts: 61
Joined: Fri Mar 18, 2016 9:45 pm

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by koedem » Wed May 01, 2019 6:15 pm

I have not looked at a thread here. Though I wouldn't be surprised if there were similar discussions in the forums here.

As for network size, yes that is true of course. I was talking about the number of nodes calculated which would be a typical measure of brute force vs knowledge. Of course you can argue that the NN is quite big and also that with 20 blocks of the NN it is in a way calculating moves there too. However that is what a human does too. The human intuition consists of pattern recognition and subconscious "calculation". Also, if you put all the patterns that a chess GM has learned in a file it sure would be bigger than 50 MB.

However I don't think that even matters so much for what the question is about. A typical human approach is to not exhaustively search a game tree but instead only look at the critical moves. And then search a few narrow lines and evaluate them based on some black box that you might call intuition or a NN in the A0 case. In that sense A0 and a strong human share more similarities than e.g. a traditional AB engine. (although a modern engine like Stockfish of course has quite a selective search as well, however the eval function compared to a human or A0 is rather "simple" there)

jp
Posts: 645
Joined: Mon Apr 23, 2018 5:54 am

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by jp » Wed May 01, 2019 10:17 pm

I'll try to post a link or quote parts. That of course discusses people's experience with Lc0, not A0, because A0 is not available.

The point (re. network size, etc.) is that nodes per second is not a complete measure of computational power required, and when it needs to use so much brute force it's hard to claim it "understands" anything. If you just store more info in a network, it doesn't mean that the program "understands" more. It's not quite the same thing, but we wouldn't think that Stockfish "understands" more if we just give it a bigger hash table, for example. A NN is not at all like a human brain. Humans keep making false analogies. There is no concept of "understanding" or "intuition" in a NN. Humans just wrongly project those features of human intelligence on to NNs. That is why (discussed in a different thread here) the NN engines generalize worse than traditional engines, i.e. they struggle with unusual or unfamiliar positions. If they really "understood" or had "intuition" in the way humans do, they would not struggle in that way. The NNs engines are clearly not discovering & remembering patterns, etc., unlike humans.

But I think the opening post is more about style of chess than the computer side. I do not think that the NN engines have a style of play like Carlsen's.

koedem
Posts: 61
Joined: Fri Mar 18, 2016 9:45 pm

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by koedem » Wed May 01, 2019 11:28 pm

I don't think humans incorrectly project "understanding" to a NN. I think it's more like we have different definitions of what to understand means.
In what sense does a human understand anything? It's just chemical reactions that are happening in the "calculation machine" that is the human brain.
The difference that you could argue for is that humans might be better at adapting things they haven't seen before, which *might* be true but even then it is debatable whether that should be called "understanding" it better.

jp
Posts: 645
Joined: Mon Apr 23, 2018 5:54 am

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by jp » Thu May 02, 2019 12:38 am

koedem wrote:
Wed May 01, 2019 11:28 pm
In what sense does a human understand anything? It's just chemical reactions that are happening in the "calculation machine" that is the human brain.
Well, that's just the mirror part of the same false belief, i.e. it's the belief that the human brain is just a machine. The question is: what is consciousness? This is a difficult question, but we don't need to answer it to say that an NN is not conscious.

If we were talking about humans, most of us would agree that those who cannot adapt their "learning" to slightly different scenarios are not really understanding. If you put the same tactical pattern on the board but add extra queens that don't make a difference (e.g. two pawns have promoted and we have four queens on the board), the NN engines can't handle it. That shows a complete lack of what we normally call understanding.

Michael Sherwin
Posts: 3024
Joined: Fri May 26, 2006 1:00 am
Location: WY, USA
Full name: Michael Sherwin

Re: Is Carlsen's approach similar to that of AlphaZero?

Post by Michael Sherwin » Thu May 02, 2019 4:48 pm

serpico wrote:
Tue Apr 30, 2019 5:59 pm
In the last few months I've read on some blogs that Carlsen's approach in this 2019 is similar to that of AlphaZero. Could there be any truth in this?
When I saw some recent games by Carlsen I did not see AlphaZero like play. I did see Bobby Fischer like play though.
I hate if statements. Pawns demand if statements. Therefore I hate pawns.

Post Reply