AlphaZero paper discussion (Mastering Go, Chess, and Shogi)

[Uberdude]

[re contempt] I don't know if other engines do anything clever like evaluating opponent's previous moves to work out they are weaker so they can get away with overplays and go for a win.

This would be hard unless you had multiple games to go on (and no chess engines that I know of have a "meta-game" like that), since by the problem setup your opponent's moves so far in this game have been about as good as your own.

Yeah, from one game would be hard, but if you could tell the bot who the opponent was and its known weaknesses or at least general level maybe. But look what I found wandering the internet:

Another approach its designers claim to use is 'opponent modeling'; Junior might play moves that are not objectively the strongest but that exploit the weaknesses of the opponent. According to Don Dailey ″It has some evaluation that can sting if it's in the right situation—that no other program has.″[2]

[Uberdude]

A few comments from a low-dan Go and somewhat better chess playing friend:

Boris, as a chess and go player can you give your impression of how much AlphaZero plays standard chess opening book lines versus going of into unchartered opening territory. I always thought human (helped by bots recently) knowledge of chess openings was closer to perfect play than for go and there was less chance to play novel early moves which weren't suboptimal (chess openings being whole board rather than just a corner like in go, and sharper). So in go AlphaGo Zero is playing 3-3 invasions before move 10 and surprising us (starting in corners and approaching with knight's moves we got right though!) and not playing many of our long josekis, is it similar in chess making innovations in the first few moves or does it do the normal book opening for the first 10 (or however many) moves (ply) and then start winning with new things after that? https://www.reddit.com/r/chess/comments ... a/dqubdxz/ suggests maybe some of the former but I don't understand chess enough.

That's a great question and I think your intuition is right. There doesn't seem to be any innovation in early moves. I'd argue that's a property of chess for being a simpler game and the ability to analyse much further than in go (all 3-3 alike moves would've been tried by strong players at some point). I've only liked through openings without a board - I'll update my answer once I've seen how it slowly grinds a win. On average the games seem to be longer than human games.

My impression of the style is that it has an extremely strong positional understanding. It evaluates and handles the most complex positions of unbalanced material very well. The way in which it keeps activity and slowly converts the superior placement of pieces in gambit situations (it sacrifices plans happily) is impressive and reminiscent of the most memorable games by top players from the past. Botvinnik vs Tal match comes to mind

[John Fairbairn]

He [Habu] also got to play with AlphaGo!

Not really - they only played 20 moves.

But I'm not sure how strong he is in Go.

Unless he's improved recently (and why would he even try?), he plays pros on 6 stones or more.

[John Fairbairn]

If you allow yourself to ponder on this, it just gets worse Kirby, I'm afraid. Huxley once said an intellectual is someone who has found something more interesting than sex. One might be tempted now, therefore, to say "sod it" and go back to prioritising sex. But David Levy, one of the big names in computer chess, recently did a late-in-life PhD at the Sorbonne on computers and sex, and this is now in book form as "Love and Sex with Robots." He even goes so far as to say that continuing advances in computers and robotics "will make legal marriages between Homo and Robo feasible by mid-century."

Best to take a detached Zen-like view, and recall that life, like go, is a game of co-existence.

[dfan]

But look what I found wandering the internet:

Another approach its designers claim to use is 'opponent modeling'; Junior might play moves that are not objectively the strongest but that exploit the weaknesses of the opponent.

Yeah, it seems I was wrong that people aren't looking at the metagame: more information here. It's difficult to implement in the current environment, though, because the standard interface for talking to a chess engine (UCI) doesn't tell it who it's playing. So any between-game hyperparameter tuning has to be done manually by an operator.

(Of course this sort of opponent modeling is totally essential for exploitative play in poker, say.)

Perhaps in the future there will be computer matches in an environment where the engines have the ability to modify themselves between games. That could be pretty cool.

[pookpooi]

Unless he's improved recently (and why would he even try?)

He's likely the best human candidate against AlphaZero, strongest Shogi professional ever, International Chess Master, Amateur dan Go player.

[johnsmith]

One thing in modern chess engines is that, to know who is better at the current position, you need to look at the estimate that shows how many pawns you are ahead. It's not percentages like in go.

For example: in all of these games Stockfish at one point thought it is 1.10 ahead, but in reality, Alpha Zero may have thought it is winning by 75% which is much more convenient. Isn't it? Maybe chess players were thinking in the wrong direction. People are saying that Alpha Zero is making a lot of long-term sacrifices.

[moha]

The chess part seems very interesting. Silver himself quoted chess and checkers as examples where search is absolutely essential (because of the tactical nature), and knowledge-based approaches are not appropriate. This is still about search of course, but I would have bet blindly that NNs are not a good deal here for their gain / slowdown ratio (I assumed 100 times, but the paper says 1000 times - and I thought the handcrafted evaluation are already reasonably good).

But in retrospect, there is some logic in this. The other day I peeked at the Houdini-Komodo match, and it seemed the current state of art is roughly 50M pos/s, for about 5G per move selection, with a typical depth or PV length of 30 plies. This would mean a searched branch factor of 2 with full minimax, so on average maybe 4 or more with perfect alphabeta (even more actually, because of selective depth). In any case, the 1000 times slowdown would be regained at a successful branch factor reduction of about 1.25 times, which is not unrealistic.

it makes me question the meaning in life for humans. If it's intelligence, then computers are beginning to have more meaning than humans.

That is still in the future. IMO intelligence means ability to solve previously unseen tasks, so I'm not sure if these examples qualify, even as early birds. Unless matchboxes are intelligent .

[Bill Spight]

One thing in modern chess engines is that, to know who is better at the current position, you need to look at the estimate that shows how many pawns you are ahead. It's not percentages like in go.

For example: in all of these games Stockfish at one point thought it is 1.10 ahead, but in reality, Alpha Zero may have thought it is winning by 75% which is much more convenient. Isn't it? Maybe chess players were thinking in the wrong direction. People are saying that Alpha Zero is making a lot of long-term sacrifices.

In traditional go evaluation, if you are ahead on points and have the move, you will win with correct play, unless there is a ko that enables your opponent to win. Otherwise, even knowing the exact point value of the position is not enough to say whether you are winning or not. That is one reason that the win rate of quasi-random rollouts was more effective for MCTS bots than the average winning point margin. In go, if you know not only the point value of the position but also the temperature, then you can make a pretty good estimate of who is ahead. For instance, if your opponent is 1½ pts. ahead, you have the move, and the temperature is 3, the game is very close; if the temperature is 1, you are a goner without a ko or the opponent's error; if the temperature is 5, you are the favorite. These temperatures might translate to an estimated win rate of less than 5% for temperature 1, 50% for temperature 3, and 66% for temperature 5. Note the non-linearity. These are my estimates of actual win rates, not win rates with quasi-random play. Monte Carlo win rates might be more like 30%, 50%, and 60%. Quien sabe?

Anyway, it may be that go playing programs that made use of estimates of both point value and temperature would perform better than programs that make use of ill defined "win rates". AFAIK, no strong program has utilized both estimates, so who knows? In any event, unlike win rates, point values and temperatures are well defined in go, even if they may be difficult to calculate in many positions.

In chess engines, point evaluations in terms of hundredths of a pawn have been very effective. OTOH, knowledge of temperature is quite crude. Is there any more accuracy than quiescent vs. non-quiescent? It is interesting that AlphaZero's use of "win rate" has proved more effective than the use of point evaluations by Stockfish. Chess seems to have no exact theory of temperature, so is pretty well stuck with point evaluations or, now, win rates, neither of which is well defined.

Edit: "Is there any more accuracy than quiescent vs. non-quiescent?" Well, yes. Checkmate and stalemate precisely as cool as you can get.

[Bill Spight]

it makes me question the meaning in life for humans. If it's intelligence, then computers are beginning to have more meaning than humans.

That is still in the future. IMO intelligence means ability to solve previously unseen tasks, so I'm not sure if these examples qualify, even as early birds. Unless matchboxes are intelligent .

Yes, I regard intelligence as the ability to do something well that you have never done before.

Fairly obviously, I think that there are different kinds of intelligence.

[moha]

IMO intelligence means ability to solve previously unseen tasks, so I'm not sure if these examples qualify, even as early birds. Unless matchboxes are intelligent .

Yes, I regard intelligence as the ability to do something well that you have never done before.

Do you know where this definition comes from? I have used it since my childhood, I don't remember the origin but I doubt I came up with it myself. A few years ago I tried to look up the source but failed.

Most contemporary definitions seem to include things like lexical knowledge and it's application ability, so very different. And there is this one too: "intelligence is what intelligence tests measure".

[Uberdude]

How about intelligence is knowing there are better things to do than join Mensa

[Bill Spight]

IMO intelligence means ability to solve previously unseen tasks, so I'm not sure if these examples qualify, even as early birds. Unless matchboxes are intelligent .

Yes, I regard intelligence as the ability to do something well that you have never done before.

Do you know where this definition comes from? I have used it since my childhood, I don't remember the origin but I doubt I came up with it myself. A few years ago I tried to look up the source but failed.

Most contemporary definitions seem to include things like lexical knowledge and it's application ability, so very different. And there is this one too: "intelligence is what intelligence tests measure".

Well, it is not exactly a definition, but I came to it when I heard a real story about someone doing something well that they had not done before. What primed me to think that that was intelligence, I don't know.

OC, I had long doubted whether IQ tests test anything, factor analysis be damned.

[Kirby]

it makes me question the meaning in life for humans. If it's intelligence, then computers are beginning to have more meaning than humans.

That is still in the future. IMO intelligence means ability to solve previously unseen tasks, so I'm not sure if these examples qualify, even as early birds. Unless matchboxes are intelligent .

Yes, I regard intelligence as the ability to do something well that you have never done before.

Fairly obviously, I think that there are different kinds of intelligence.

@moha: Nice video. That matchbox experiment is pretty cool.

@Bill: I'd say that I'm not really intelligent at anything, then. I rarely do something well the first time I try it. Only after practice can I get any sort of competency.

Best to take a detached Zen-like view, and recall that life, like go, is a game of co-existence.

Good advice, John.

An often overlooked implication of rooting the meaning of human life or the moral status of humans in our special intellectual capacities is that human beings lacking them have less meaningful (or non-meaningful) lives or do not have the moral status of unimpaired human beings. This is an implication that I certainly find problematic, and so it tends to undermine the idea that we should base our understanding of these concepts in the special intellectual capacities of human beings.

You bring up a good point, Monadology. I also find it problematic to equate moral status with particular intellectual capacities, even for unimpaired human beings. My perspective was more reflective in terms of finding self-worth, perhaps (e.g. what makes me valuable as a human being?). But maybe the implications are similar. If I try to equate my value as a human being, I suppose there is still the aspect of being a father to my children, and a husband to my wife. I'd say that I bring some value from that dimension, which can't currently be replaced by machines. Maybe that status could be replaced by another person, but I hope it won't be.

Maybe people are like go stones - no inherent value, but useful at the right place and the right time, under the right circumstances, working in coordination with the rest of the universe...

Or maybe I'm just pushing an analogy too far :-p

[sorin]

I'm wondering why DeepMind choose Shogi out of many board games beside Go

Another usual suspect: Xiangqi ; ...Chaturanga.

For diplomatic reason, choosing Xianqi to conquer by (non-Chinese) AI is a really bad choice.

Really? How about Go then?