Google reveals secret test of AI bot to beat top Go players • Hacker News Board

Re: Google reveals secret test of AI bot to beat top Go players

#201

Post by taejo » Fri, Jan 06, 2017, 9:12 AM UTC

gertef wrote: post #28

bluetwo wrote: post #21

Earlier quoted context omitted.

I don't know much about Go but I thought a comment about AlphaGo was interesting that it played in a way to marginally beat the player, which was different that most masters played, which is to clearly beat the opponent by as wide a margin as possible. Is this accurate? Does MasterP also use this style? Are there humans that can play this way? (I'm asking you because you seem to know what you are talking about here.)

Humans play for a large lead because they don't have enough memory/power to accurately estimate the value of their positions, so they play for a buffer -- AlphaGo has higher confidence in its valuation, so it can play it closer -- ~85% confidence of winning by 5 stones (with room for error) vs 99% chance of winning by 2 stones

Another factor is that many micro-endgame sequences simply have a "correct answer" that loses the least points. Any human who's played Go for more than a few months knows these sequences, and if they choose to answer a certain move, will prefer the answer which is "always correct" to another move which would also win the game. This naturally leads to the winning player preserving their margin even when they could throw it away, while the machine has no such bias, and will just as happily throw away the margin as preserve it.

Re: Google reveals secret test of AI bot to beat top Go players

#202

Post by 21 » Fri, Jan 06, 2017, 10:23 AM UTC

daveguy wrote: post #173

21 wrote: post #120

Earlier quoted context omitted.

It would be super-weird if AlphaGo became worse that the previous version.

Neural networks have a bad habit of regressing. When training continues they tend to replace earlier skill with newer skill.

The team would surely detect that. In the original paper they even mention this issue, and as a safe-guard they also test later networks against previous versions.

Re: Google reveals secret test of AI bot to beat top Go players

#203

Post by ousta » Fri, Jan 06, 2017, 10:42 AM UTC

stouset wrote: post #78

dfan wrote: post #7

One thing that isn't made clear in this writeup is that Master plays in a very nonhuman style, as opposed to the version of AlphaGo that beat Lee Sedol, which mostly played like a strong human except for a few surprising moves. My first guess when I saw Master's games was that it was a program like AlphaGo that had its policy network trained from scratch rather than being bootstrapped by being given the goal of imita…

As computers are able to evaluate positions faster (and therefore deeper), the "godlike" tactics are dominating over human-style strategy. It used to be that computers played "computer-like" moves because they didn't understand the position. Now, they play computer-like moves because "understanding" the position isn't as important as just being able to see 25+ moves ahead. In a nutshell, positional play in chess is s…

if we had unlimited computing power, couldn't we solve any know problem in a way that will ALWAYS outbeat any other way to solve it?

Re: Google reveals secret test of AI bot to beat top Go players

#204

Post by daveguy » Fri, Jan 06, 2017, 12:57 PM UTC

21 wrote: post #202

daveguy wrote: post #173

Earlier quoted context omitted.

Neural networks have a bad habit of regressing. When training continues they tend to replace earlier skill with newer skill.

The team would surely detect that. In the original paper they even mention this issue, and as a safe-guard they also test later networks against previous versions.

True, they are aware and monitor for it. Just pointing out that it's not a super weird thing to happen with a NN. In fact it's a super normal thing that has to be continuously monitored.

Re: Google reveals secret test of AI bot to beat top Go players

#205

Post by andorov » Fri, Jan 06, 2017, 3:01 PM UTC

SAI_Peregrinus wrote: post #194

andorov wrote: post #192

Earlier quoted context omitted.

But no one could play because you can't create a Turing machine that provably halts, no?

You can prove that any given Turing machine halts or does not halt, but there's no single algorithm that can prove that for every Turing machine. The trivial case of a Turing machine that can be proven to halt is one with only one state: halted.

Yes but this case requires Turing machines that produce very very large (BB large) outputs and then halt. Is that provable for a specific machine?

Re: Google reveals secret test of AI bot to beat top Go players

#206

Post by hyperpape » Fri, Jan 06, 2017, 4:08 PM UTC

stouset wrote: post #199

hyperpape wrote: post #183

Earlier quoted context omitted.

I would not describe the moves dfan is talking about as tactical. Rather, they appear to exhibit a very different positional judgment from what humans use/different strategic aims. What you say sounds plausible, but I do not believe it is backed up by any analysis of the games.

Positional play is nothing more than our human attempt to perform short-circuit evaluation of positions by using heuristics that are easier than calculating non-forced lines precisely to 20+ ply. If you imagine a perfect chess-playing computer, it would have no need of any sort of positional evaluation — literally every move it chooses would be based upon the pure tactical outcome of having evaluated every subsequent…

dfan is talking about go. What you're saying is accurate about chess, but go programs aren't the same as chess programs.

Re: Google reveals secret test of AI bot to beat top Go players

#207

Post by skmurphy » Sat, Jan 07, 2017, 12:25 AM UTC

Deimorz wrote: post #140

daveguy wrote: post #119

Earlier quoted context omitted.

Centaur chess players do not outperform the best chess engines. That is based on a misrepresentation of one of the first centaur chess tournaments. Human modified moves only reduce the strength of the best chess engines now. The benefit that was reported was one of tactics one centaur chess competitor had over other centaur chess competitors , not centaur chess vs stand alone chess engine. It is a misunderstanding th…

And it's such a strange claim, you'd think it would be obvious that it makes no sense. Nobody would ever try to claim "I think that a team of a top pro and a mediocre amateur would be stronger than the top pro alone", but that's basically what they're doing with centaur chess.

Why do startups staffed by amateurs sometimes outperform teams of experts at established firms? Creativity and new perspectives can yield surprising results.

Re: Google reveals secret test of AI bot to beat top Go players

#208

Post by monktastic1 » Sat, Jan 07, 2017, 1:38 AM UTC

fryguy wrote: post #123

btilly wrote: post #81

Earlier quoted context omitted.

How about the following statement? "There is no widely played game for which an Elo rating system shows a larger measurable range of skill than Go." This is a precise and measurable statement. The standard deviation of the Elo system used in Chess is 200, and the range from the best humans to rank amateurs is 14 times this standard deviation. When the Elo system was adapted to Go, the standard deviation was set to 10…

Complex is being used in a different context here though. You are meaning complex as in strategic complexity. The parent (I think) is using it in terms of the games rules. Go is pretty simple in principle: place stones on the board one at a time, and remove enemy pieces completely surrounded by your own. But it's fairly complex to know what exactly constitutes being surrounded. Chess has 6 different kinds of pieces t…

Know of any resources that help figure out what the Elo SD of a particular game is, and the range? Also, know of anything written about various ways of measuring the complexity of a game? I'd never thought about the number of SDs thing.

Re: Google reveals secret test of AI bot to beat top Go players

#209

Post by Radim » Sat, Jan 07, 2017, 6:27 AM UTC

hyperpape wrote: post #184

Radim wrote: post #154

Earlier quoted context omitted.

You mean, like these three AlphaGo self-play games from September? ;) https://deepmind.com/research/alphago/alphago-games-english/ (analysis by Gu Li and Zhou Ruiyang, two top pros; standard komi)

Note that these games look much more human than the ones dfan was describing. There are surprising ideas, but they are still much more normal.

I don't know about that.

I'm a mere 1 dan, but I fail to see this marked "difference in normality".

What specifically looks "much less human" to you in these newer games?

Re: Google reveals secret test of AI bot to beat top Go players

#210

Post by drdeca » Sat, Jan 07, 2017, 7:44 AM UTC

nightcracker wrote: post #170

kenshaw wrote: post #136

Earlier quoted context omitted.

Probably. I think it basically proves that any rule structured game without randomness is solvable.

There is a trivial disproof of your claim. Two players write a turing machine with at most n states with two symbols. The player that produced a terminating turing machine that produces the most 1 symbols on the tape before terminating wins. The optimal strategy for this game is producing a busy beaver, a feat shown not to be computable.

this comment im making mainly me thinking out loud. It is not insightful:

I was thinking that for a fixed n that doesn't really work, because there are only finitely many options, but I guess if n>~2000 , ZFC cannot show the winning strategy to be the winning strategy? Is that what you meant?

Given any two machines which halt, finding the one that ends with more ones is computable. Assuming at least one of the two machines halts, which one wins can be computed in the limit? By which I mean, if the process is allowed to have a "who is currently winning" (the one that already halted if only one has, the one that halted later if both have, or neither if both haven't) thing, and the limit of that is whatever it eventually never switches from. I guess that works even if neither halt.

Uh... I'm just saying stuff you already know to try to think through it myself.

Edit: I guess the question is then, what exactly do we mean by solvable?

Do we mean that there is an algorithm that outputs an optimal move on every turn? For any n, there is such an algorithm. The one that has the correct move hard-coded. Maybe we mean that there is an algorithm that probably always outputs an optimal move? In this case, well, I suppose it depends on the axiom system. Hm.