Alpha Zero’s “Alien” Chess Shows the Power, and the Peculiarity, of AI | Hacker News Reader

Alpha Zero’s “Alien” Chess Shows the Power, and the Peculiarity, of AI

technologyreview.com

26 pointsrbanffy11 comments

Its only alien because for decades people sculpted chess to look like accounting: AlphaZero doesn't waste time evaluating advantage in centipawns for each position(which is unreliable long-term), because it has more general evaluation function that is not as dependent on search algorithm as alpha-beta. That evaluation function is just hidden behind the neural network weights and is probably much faster than AlphaZero itself(a generic network trainer): essentially behind it is a series of rules and heuristics that would translate to fast classical algorithm.

You could argue that behind the brain of a chess master is also a “series of rules and heuristics”, but as this series may be infinitely long and isn’t feasibly determinable it’s difficult to say in what sense that’s actually true. Further, it’s likdly that the series of mostly-linear transforms in AlphaZero (or, the series of neurochemical transforms in a chessmaster’s brain) is the only form in which these rules would be efficiently computable.

One way of looking at the difference between AlphaZero and traditional game AI is that the former uses irreducibly conplex heuristics (as far as we know,) just as we think brains are capable of.

> He pointed, among other examples, to the intelligence of Hassabis and his colleagues in devising, designing, and building the program in the first place. “That’s almost as impressive as a queen in the corner,” he quipped.

That sentiment that pops up whenever machines surpass humans is quite silly to me.

As if they were discussing a crane and say: This crane can lift many tonnes of weight more than any human can but we must remember that it was put together from parts by strong humans, strong humans designed it and operate it, said body building expert.

Who cares? Machines are stronger than you, calculate better than you, play games better than you, often design better than you. And people react as if they just found out that they have bigger d*cks than them and last longer. Sure they do but isn't it juvenile to feel insecure and defensive each damn time?

Turns out that really is a concern for a lot of people.

http://www.motherjones.com/politics/2016/04/ted-cruz-dildo-b...

AlphaGo isn't a human, but it played Go in a human manner. When reviewing the games it is hard for Go players not to talk about 'intent' and 'planning'. The actual process for making the move was actually very human, in my view. Superhuman execution, but the basic process was how a human played.

It is interesting to see how that translates into Chess. When AlphaZero wins it is playing games that exercise a beautiful positional judgment. There are a couple of very elegant games where a material advantage for Stockfish just didn't translate into a meaningful edge on the board.

If the Go experience is any guide, this is a glimpse in the direction of the ideal we should aspire to playing Chess.

> When reviewing the games it is hard for Go players not to talk about 'intent' and 'planning'.

An interesting point. Of course the AI is not planning with intent, but the language is useful. But it IS seeing that if it does X, then the opponent is likely to do Y, which will allow it to do Z. And note that when we see a human player playing X, we say he/she is planning Z, but in reality that's not quite true. What's happening is just the same: he/she is expecting Y, which will allow Z. The process is not so different.

In the spectrum between "continuous" and "discrete", I wonder where the chess problem lies. AlphaZero learning from zero hard coded behaviors, and following the description of exploring all the paths one might put it close to being continuous. The thing is, AlphaZero shines by aggressively not exploring lots of paths. So there must be some kind of abstraction it relies on to filter theses paths; and this represents some sort of local minima of the topology of the problem.

I'd be very interested in knowing if some rules can emerge from this. And if it can be extracted from the AlphaZero internal model. Hopefully rules that can be interpreted by humans. It could be measured by: what percentage of play decisions can be reproduced with a rule expressed with under 150 symbols? Both for chess moves themselves, and for a path-pruning strategy.

I think in the future explainable AI will be a standard part of the AI/ML stacks. The DARPA launched recently a program titled just like this [1]. They talk about deep learning; but I'd be very interested if some results can be had with Reinforcement Learning.

[1] https://www.darpa.mil/program/explainable-artificial-intelli...

>Of course the AI is not planning with intent

Isn't it, though? It has an intuition built up from the memory of millions of games it's played before. It knows that one step now leads to a good situation later. What other definition of "intent" can you use where AZ's playstyle doesn't fit the criteria?

And does having intent even mean anything when the goal is to win? If I intend to pursue a certain specific action plan, is that somehow better or more appropriate to the game than playing with infinite flexibility wrt the possible trajectories of the game through state space?

That "intuition" is basically guessing. Machine learning is just a new form if brute force

Only in the sense that anything at all is just a form of brute force.

> And does having intent even mean anything when the goal is to win?

I think "intent" implies consciousness (and so, when we say that a spider intends something, we are using the term metaphorically), but probably this is one of the many cases in AI where it's not worth fighting over the definition. Can a machine be conscious, can it have "real" intent, etc -- can a submarine swim.

Reply on news.ycombinator.com