AlphaGo Retiring Only From The Public, Not From Research

[Bonobo]

https://www.theverge.com/2017/5/30/15712300/alphago-ai-humanity-google-artificial-intelligence-ke-jie

(Emphasis added)

I asked Hassabis whether he thought this system could work without the initial dataset taken from historical games of Go. “We’re running those tests at the moment and we’re pretty confident, actually,” he said. “The initial results have been that it’s looking pretty good. That’ll be part of this future paper that we’re going to publish, so we’re not talking about that at the moment, but it’s looking promising. The whole idea is to reduce your reliance on that human bootstrapping step.”

[EdLee]

Tengen or not tengen, that is a question.

[Bonobo]

Tengen or not tengen, that is a question.

I bet And … probably (loosely? tightly?) connected, the “proper” value of Komi, I’d think.

[Kirby]

Not exactly related to the OP, but I thought this part of the article was interesting, where they were describing how they addressed bugs found from the version that lost game 4 against Lee Sedol:

Another thing we did is when we assessed what kinds of positions we thought AlphaGo had a problem with, we looked at the self-play games and we identified games algorithmically — we wrote another algorithm to look at all those games and identify places where AlphaGo seemed to have this kind of problem. So we have a library of those sorts of positions, and we can test our new systems not only against each other in the self-play but against this database of known problematic positions, so then we could quantify the improvement against that.

I often like to consider how I can improve my own game by observing what AlphaGo has done to improve. In this case, it might seem analogous to looking through my past games, finding trends in positions that I screw up, and then perhaps generating a set of problems related to that theme.

Sounds like a lot of work to do manually.

Maybe it'd be possible to use one of those programs that have an open source evaluation function. I could run through old games and pick moves I played where the winning probability dropped significantly. That'd give me a set of positions to look at. Maybe I could draw inferences from that.

[pookpooi]

Sounds like they didn't make the 'learning from zero' version strong enough just in time for the match and decide to use the ordinary version instead.

[hydrogenpi7]

Sounds like they didn't make the 'learning from zero' version strong enough just in time for the match and decide to use the ordinary version instead.

Also if Deepmind was fairly confident they could give a top pro 2 or 3 stones handicap and still win, I think they would have done it this time, or announced in a while that this new event would occur. But maybe they ran into a hard wall and some insurmountable roadblock that mathematically proved to them they it will never in the future be possible to give a top pro more than 2 stones and still win for AI Go, then that could help explain why the sudden retirement after achieving the "pinnacle".

[EdLee]

Also if Deepmind was fairly confident they could give a top pro 2 or 3 stones handicap and still win, I think they would have done it this time

Not necessarily; more likely a no.

[Mike Novack]

We can conclude little about in what directions they will go.

The problem is that we are seeing this "managing to train a neural net to play top level go" and what we are hearing about now, finding a way to do that without starting with a human database for them hs ANYTHING much to do with go. Other than the fact that they used THIS problem (go) to learn about tackling difficult problems of this sort.

There is no money in it. "Advertising" value yes, but that they now have gotten about all the bang for the buck that can be had << being able to give a top pro three stones would add little outside of the relatively small world of those who play go >>

Think of the HARD problems that would have real profit potential. For example, "self driving cars" under the conditions that all other vehicles and pedestrians and animals will strictly obey the rules of the road. We are probably there. But the MUCH harder problem (that we human drivers have to solve all the time) of judging when some other vehicle (or pedestrian or animal) is likely to violate one of those rules, that's another matter. Not straight forward, as I doubt many of us could explain HOW we know/sense, what are the cues we pick up ro decide that other driver who should yield will, .... or won't.

[pookpooi]

From article
"In fact, AlphaGo Master uses much less power than the version of AlphaGo that beat Lee Se-dol; it runs on a single second-gen Tensor Processing Unit machine in the Google Cloud, whereas the previous version used 50 TPUs at once."

David silver added from this video that single machine contains 4 TPU. So this is what 'roughly tenth time power from last year version' quote originated from. Kudo to Chinese reporters that ask this question so many time we finally have exact answer from DeepMind. Still, many questions left...

[moha]

Also if Deepmind was fairly confident they could give a top pro 2 or 3 stones handicap and still win, I think they would have done it this time, or announced in a while that this new event would occur. But maybe they ran into a hard wall and some insurmountable roadblock that mathematically proved to them they it will never in the future be possible to give a top pro more than 2 stones and still win for AI Go

But how would you explain AlphaGo's results, unless it's at least a few stones stronger than pros? In about 70 games it only had one loss (and even that loss is a bit doubtful). This do sound like more than 2 stones...

[zermelo]

But how would you explain AlphaGo's results, unless it's at least a few stones stronger than pros? In about 70 games it only had one loss (and even that loss is a bit doubtful). This do sound like more than 2 stones...

I think it's quite obvious that when you approach perfect play, ratio (difference in winning %) / (difference measured as handicap stones) goes to infinity. The problem is that we don't know anything about how far top pros are from perfect play (in handicap stones).

And the other explanation, that many have mentioned: maybe alphago, just because of training history or whatever, cannot handle handicap play well. With 2 stones on board, maybe it's evaluation says that it has 10 % winning probability, and it starts doing stupid things. OK, it can play handicap games against other alphago versions. But as they said, it learns to attack specific weaknesses, so maybe it learns some specific semeai/l&D that are the blind spot of the earlier version, and learns to drive the game to that. But human pros may not have those same blind spots at all.

[moha]

I think it's quite obvious that when you approach perfect play, ratio (difference in winning %) / (difference measured as handicap stones) goes to infinity. The problem is that we don't know anything about how far top pros are from perfect play (in handicap stones).

The first statement is not as obvious as it seems. How would you define being 1 stones away from perfect play? Probably like it wins 50% with 1 extra handi stone (assume white always gets komi for correctness). Without handi, perfect play against such a player may get a winrate higher than 2/3, but still not 100% IMO. OC it won't lose, but there will be some ties.

Also, even if the strength function is significantly different at that level, such differences may quickly disappear, like when this "-1" stones strength player plays with a "-2" player. And I'm pretty sure AlphaGo is still quite a few stones from flawless play.

MC not being well suited for handi can be a more realistic problem, but sooner or later this will be solved as well. If not with the retiring AlphaGo, then it's successors. We need to know how far pros are from perfect play, afterall...

[moha]

when you approach perfect play, ratio (difference in winning %) / (difference measured as handicap stones) goes to infinity

Actually I think there is more to this. We like to measure strength with a single number for simplicity. But instead of "- stones" (extra stones needed for 50% winrate against perfect play) let's use "- points" (points dropped on average during a game, roughly the extra komi needed for 50% winrate - may loosely correspond to stones when divided by ~14).

And add a variance/deviation factor. So perfect play is "0,0" (drops 0 points with no variance). The "-1 stones" player is, say "-15,10". So he sometimes drops only 5 points instead of 15, which would increase his winrate elsewhere. But against perfect play -5 is still loss. He also loses the chance to play significantly above his average. However, a similar thing can happen at much weaker levels, if we imagine an opponent with no variance, such as "-100,0" (a very special amateur dan? ).

Half joke OC, although variance can explain some subtleties, there are limits for how well results in even games can be translated to handicap stones. Also points dropped against perfect play may not correspond well to points dropped against weaker opponents.