23 November 2018

A Conversation with Demis Hassabis

Continuing a series of posts on AI/NN, I'll suspend the nuts-and-bolts overview seen in the previous post, GPU Benchmarks, to cover a rare chess-related appearance of AlphaZero's guiding light, Demis Hassabis, who visited the 2018 Carlsen - Caruana venue. Later he tweeted, A real honour to make the ceremonial first move of match 8 of the World Chess Championships today... (twitter.com). After making the first move, he visited the commentary room.

World Chess Championship 2018, Day 8, First moves
(youtube.com; 'Streamed live on Nov 19, 2018')

At around 23:40 into the Youtube clip, Hassabis joined the official commentators for the match. Left to right: Judit Polgar (JP), Anna Rudolf (AR), Demis Hassabis (DH).

AR: We are so glad that Mr. Demis Hassabis has returned to our studio. Welcome back, Demis. The co-founder and CEO of Deepmind as we discussed already and if you were not here yesterday I don't what you were doing because you missed a very insightful conversation with one of the brightest minds in the world. We started out with AlphaGo and AlphaGo Zero and today we will discuss further topics about artificial intelligence and chess as well.

DH: It's great to be here and great to do the first move, because it was really interesting feeling the intensity in the room. It's quite amazing the amount of energy in the room. You go inside and you see they're superconcentrating. It was quite an interesting experience on that side of the glass. It's almost claustrophobic inside. It feels like the room is not big enough to contain the energy of the two players.

AR: Did you expect 1.e4 or did you think Fabiano would tell you something else?

DH: I wasn't sure what he was going to do. I'm pleased to see the Sveshnikov because I played this for Black. It was quite a coincidence it was this match [game?] I got to do the first move. We'll see how this pans out.

AR: We will definitely ask for your expert advice. As mentioned we are going to discuss further topics about artificial intelligence. Yesterday we were so sorry when we had to say goodbye to you. We thought that there are so many other fascinating topics about artificial intelligence that we wanted to have you here for hours. Thank you so much for coming back. The first topic I wanted to discuss was how differently AlphaZero is thinking about the game of chess -- if we compare it to humans and if we compare it to computer engines.

DH: There are two interesting things to say about that. One is how many moves do the chess engines calculate per decision. Human grandmasters maybe look at 100 moves, something like that order [of magnitude], to be able to make a decision. Something like Stockfish and traditional chess engines, they look at 10.000.000 moves before they make a decision. AlphaZero is somewhere in between, so it looks at 10.000 moves before it makes a decision. It's not as efficient as human decision making but it's much more efficient than traditional engines. It looks at a lot less moves because it's better at evaluating positions.

The second thing that's interesting about it is because it doesn't have in-built moves. It doesn't have 'a Queen is nine points, a Rook is five points'. It doesn't know anything about those piece values, so it senses everything in the context of the current position. We speculate that it's much easier for it to make long term sacrifices; for example, because it doesn't have to overcome its in-built programming. Say it's going to sacrifice an exchange. An engine like Stockfish would have to calculate that it's going to get enough in return for that two points difference. Whereas AlphaZero doesn't have that rule in-built so it can just decide the Rook is an asset, the Knight is an asset, and in the current position that Knight is a stronger asset for the opponent than my Rook. It can make the sacrifice even if it can't calculate explicity that it's going to get enough compensation. It can just sort of decide contextually that in these kinds of positions that exchange is worthwhile.

JP: How does he make a difference between Knight and Rook if he doesn't know that the Rook is worth more?

DH: I guess it's learned over the millions of games playing against itself that the Knight will give a better outcome over the whole experience. In the context of this particular position it can make a decision to make the sacrifice. Stockfish can do that, too, but it would need to calculate quite carefully that it's going to get enough in return to overcome this in-built rule that you are losing two points.

JP: Does AlphaZero have intuition?

DH: It is sort of like a very intuitive player. It does it more by feel in effect because it's taking the pattern of the current board and deciding that this is worthwhile. It's not necessarily explicitly calculating out. It's more akin to something like intuition in human terms. Of course, it doesn't know anything about intuition or any of these terms we're using -- it's just a computer -- but it's more like that and it comes out in the style of the play. It really likes sacrifices. It's very positional, I would say.

JP: How many moves ahead does it calculate?

DH: It can calculate quite deep lines if it needs to, but it only looks at 10.000 moves per decision. It's quite a lot compared to a human player, but it's much, much less than you're used to with a chess engine which is millions of moves. In order to compensate for that lesser amount of caculation it has to have better evaluation.

JP: What makes the difference about how deeply it goes into certain lines? We have strategic positional lines and we have tactical. Does it understand that in a tactical line you really have to go all the way?

DH: Not explicitly. Chess engines have this extra calculation. They know that if it's an imbalanced position then they should calculate more : there are special moves for that as well. AlphaZero doesn't have anything explicit about that. If it feels that a certain line is unresolved -- it doesn't quite know what's happening -- it can search in more depth. So there is some implicit way it does that, but not in the explicit way that chess engines do. This is something very interesting for us. We've only just built the system, so the next stage over the next year is to try and reverse engineer it to see how it's making its decisions. At the moment it's more like a black box. It makes these decisions, but I would like to know, for example, what does it rate a Rook and a Knight overall. We don't actually know. It doen't express its evaluation in terms of Pawns, like chess engines do. It expresses it in terms of percentage chance of winning.

At this point, AR discussed the games played during the AlphaZero - Stockfish match. See The Constellation of AlphaZero (December 2017), for the earliest posts on this blog about the match.

AR: AlphaZero not only won the match very convincingly, but it came up with these sacrifices that you mention, not just the positional sacrifices we are used to in human games, which would be a Pawn sacrifice, but it went on sacrificing a Bishop, a Knight, for something very long term, not a calculation that there would be a reward for the sacrifice 10 or 20 moves later. Then there was another move I really loved, a Queen move. I felt like we are learning something from this AI bacause that's neither a human move nor an engine move.

DH: Exactly. There were a lot of examples. I think people were surprised that it was making very unusual moves that were kind of alien. They weren't really the kind of moves a computer engine would do. I hope that's going to give strong chess players new ideas, maybe usher in a new era of creativity, because it's a very interesting style.

We're going to release a lot more games and then people can see even further what this style is. One thing it really favors is mobility. It really likes mobility and optionality for its pieces, and it likes restricting the mobility of the opponent, including using Rooks, especially Rooks, on outposts, very advanced outposts, which is quite unusual for chess.

Here the discussion turned to the match between AlphaGo and Lee Sedol, where the Korean player learned from AlphaGo's play. Could anything be learned about chess from AlphaZero?

DH: We had a couple of very strong chess players come in and look at the games and help us analyze them. One thing they told me, that stuck in my mind, is that it felt to them as if the board was much bigger somehow. I thing people will see that when they see the games. It plays on one wing, creates a few weaknesses, then it switches all of its pieces to another wing and makes more weaknesses. Then it finally goes back and the opponent's position collapses because there are a few too many weaknesses. It's very interesting how it controls that situation. People are going to find these newer games, with an even stronger version of AlphaZero quite fascinating.

AR: I believe it was former World Champion Garry Kasparov who said that IBM's Deep Blue basically caused the end of an era but AlphaZero is the beginning.

DH: I hope so. It was very kind of him to say so. Garry has spent a lot of time thinking about computer chess and he was right in the middle of the biggest moment of all. It's been fascinating talking to him about that. I think what he meant was that he realized that we're building these general systems not only just to learn all types of different games, including chess, but eventually to apply them to real world problems in science and medicine. These AlphaZero techniques are not built specifically for chess -- it just learns it for itself -- these techniques can be used for other complex domains in the real world.

After further discussion about how each version of AlphaZero learns from its previous version and what the Elo rating of the AI engine might be, the conversation turned back to the Caruana - Carlsen game. For another video showing Hassabis, see Kasparov Talks at Google (June 2017).

No comments: