AlphaGo's evaluation function

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Kappatoo
Posts: 15
Joined: Sat Sep 23, 2017 8:33 pm

AlphaGo's evaluation function

Post by Kappatoo » Sun Nov 26, 2017 8:42 pm

AlphaGo Zero's neural network alone, i.e. not doing any search, is only marginally weaker than the AlphaGo version that beat Fan Hui. AlphaGo Zero's evaluation function is thus on the level of a professional Go player. From a chess player's perspective, knowing very little about Go, this strikes me as remarkable.
In any case, it suggests to me that human Go players don't calculate nearly as much as chess players. Is that a reasonable assumption? And does anyone have a rough estimate how strong elite Go players would be if they didn't do any search, just playing based on their intuition?
Also, do you have an idea how strong top chess engines would be without any search?

Pio
Posts: 119
Joined: Sat Feb 25, 2012 9:42 pm
Location: Stockholm
Contact:

Re: AlphaGo's evaluation function

Post by Pio » Sun Nov 26, 2017 9:56 pm

How do you know that the network does not do a shallow search ;)

Actually you do need to search at all if you have a perfect evaluation. And even better you only have to know what move is the best at every position.

I think you could do really cool things with a neural network in chess and I have some crazy ideas that might work :)

Uri Blass
Posts: 8591
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: AlphaGo's evaluation function

Post by Uri Blass » Sun Nov 26, 2017 10:02 pm

Kappatoo wrote:AlphaGo Zero's neural network alone, i.e. not doing any search, is only marginally weaker than the AlphaGo version that beat Fan Hui. AlphaGo Zero's evaluation function is thus on the level of a professional Go player. From a chess player's perspective, knowing very little about Go, this strikes me as remarkable.
In any case, it suggests to me that human Go players don't calculate nearly as much as chess players. Is that a reasonable assumption? And does anyone have a rough estimate how strong elite Go players would be if they didn't do any search, just playing based on their intuition?
Also, do you have an idea how strong top chess engines would be without any search?
I wonder if humans can play a the same level of alphago Zero's evaluation without doing search simply by learning alphago zero's evaluation without search and calculating it or maybe the calculation of this evaluation is something that is going to take humans many hours for a single position.

Kappatoo
Posts: 15
Joined: Sat Sep 23, 2017 8:33 pm

Re: AlphaGo's evaluation function

Post by Kappatoo » Mon Nov 27, 2017 9:25 am

Suppose someone managed to determine my (chess) evaluation function and translate it into code. I doubt that I could learn even that code by heart, let alone apply it in a game. And I would assume that AlphaGo's evaluation function is more sophisticated than that.
That said, I am curious what the approximate size if AlphaGo's evaluation function is.

Werewolf
Posts: 1197
Joined: Thu Sep 18, 2008 8:24 pm

Re: AlphaGo's evaluation function

Post by Werewolf » Mon Nov 27, 2017 10:47 am

Uri Blass wrote:
Kappatoo wrote:AlphaGo Zero's neural network alone, i.e. not doing any search, is only marginally weaker than the AlphaGo version that beat Fan Hui. AlphaGo Zero's evaluation function is thus on the level of a professional Go player. From a chess player's perspective, knowing very little about Go, this strikes me as remarkable.
In any case, it suggests to me that human Go players don't calculate nearly as much as chess players. Is that a reasonable assumption? And does anyone have a rough estimate how strong elite Go players would be if they didn't do any search, just playing based on their intuition?
Also, do you have an idea how strong top chess engines would be without any search?
I wonder if humans can play a the same level of alphago Zero's evaluation without doing search simply by learning alphago zero's evaluation without search and calculating it or maybe the calculation of this evaluation is something that is going to take humans many hours for a single position.
I was wondering this: In chess a human can play at 2800 elo by searching at 1 or 2 nps! That's incredible.

If a NN chess machine got as good an evaluation function as a GM, but searched 1 million nps (should be achievable)....

PK
Posts: 826
Joined: Mon Jan 15, 2007 10:23 am
Location: Warsza
Contact:

Re: AlphaGo's evaluation function

Post by PK » Mon Nov 27, 2017 12:41 pm

In Go, line between search and evaluation is somewhat blurred. You can calculate that certain shape (configuration of stones) cannot be cut (stones, or most of them, cannot be separated from each other given correct play) or you can know it without calculation. The more shapes you know, the more resources you have to calculate something beyond basic stuff.

Also (being 6kyu myself and seeing some amateur dan games) I can say that Go players can calculate longer lines than chess players. This is because a go move changes position much less than a chess move. In some local situations seeing 10 plies ahead is nothing to be proud of, and can be done in a matter of seconds (most of these situations would not have to be calculated at all by a dan player).

Kappatoo
Posts: 15
Joined: Sat Sep 23, 2017 8:33 pm

Re: AlphaGo's evaluation function

Post by Kappatoo » Mon Nov 27, 2017 2:54 pm

I was wondering this: In chess a human can play at 2800 elo by searching at 1 or 2 nps! That's incredible.
This leads to another question I had: How strong would top engines be that calculate as fast as humans?

Kappatoo
Posts: 15
Joined: Sat Sep 23, 2017 8:33 pm

Re: AlphaGo's evaluation function

Post by Kappatoo » Mon Nov 27, 2017 2:59 pm

In some local situations seeing 10 plies ahead is nothing to be proud of
Do you think this is because the patterns are common and thus familiar to strong players? (Comparable to calculating a smothered mate sequence, or a bishop sacrifice on h7.)
It doesn't seem to me that this is so different in chess, by the way. Many forced lines are easy to calculate to more than 10 plies.

clumma
Posts: 177
Joined: Fri Oct 10, 2014 8:05 pm
Location: Berkeley, CA

Re: AlphaGo's evaluation function

Post by clumma » Mon Nov 27, 2017 7:25 pm

On a related note, I was watching Carlsen play blitz and bullet on chess.com the other day. His Elo advantage against other super-GMs increases at each faster time control, where calculation plays less and less a role.

In interviews, he has repeatedly stated that his "intuition", and not calculation ability, is the primary source of his playing strength. I always took this with a grain of salt (top-performing people in any field are seldom able to explain their process) but now I am beginning to take it more seriously.

-Carl

User avatar
fern
Posts: 8755
Joined: Sun Feb 26, 2006 3:07 pm

Re: AlphaGo's evaluation function

Post by fern » Mon Nov 27, 2017 8:08 pm

My guess is that evaluation and search are NOT different things that can be treated as isolated. In the first place you can know that a move is the best in certain position because at least ONCE the search was done and alouded to get that conclusion.
How you can evaluate without knowing the result and how you know the result without a search?

Fern

Post Reply