Human level far below the optimum

[bayu]

I got intrigued by the fast improvement of bots.

In chess, it took 10 years to improve bots who won against top humans to bots who never lose to humans. In Go, this took less than a year (or about a year when you wait for the announced games with longer time settings.) Even if you add another year because somebody might eek out a win during 2017, the time span is remarkably shorter than in chess.

When something strives for an optimum, progress is steep at the beginning and flattens when reaching close to optimum level. (like -y=a^(-x)+1 with (a>1), if this helps visualising it). Of course it's a rough ride with jumps. The jumps being breakthroughs of any kind. Monte Carlo was one. Using the neural nets and architecting AlphaGo was one. But since AlphaGo played Fan Hui, the progress seems to be of the steady kind with only small breakthroughs. The version now probably works similarly to the one we saw in the games against Lee Sedol and Fan Hui. And seeing how fast it went from Fan Hui-level to Lee Sedol-level it's not slowing down with Master (well it is hard to measure, maybe it slowed down. But as Ke Jie changed his opinion from "AlphaGO can't beat me" to "it's too strong", I don't think it did slow down).

In conclusion, computer go is nowhere close to optimal play. And looking at mortals, they too. Are chess players closer to the optimal play than go players? Or had the leasurely pace of computer chess progress more to do with hardware restrictions and therefore resulted in a shallower curve?

[pookpooi]

It can be answered in resource aspect. IBM immediately abandon computer chess right after it won against Kasparov in 1997. While DeepMind increasing AlphaGo team members even after Lee Sedol match. If DeepMind doesn't publish any paper, and retire AlphaGo immediately in March, it gonna takes more time for sure.

[oren]

It can be answered in resource aspect. IBM immediately abandon computer chess right after it won against Kasparov in 1997. While DeepMind increasing AlphaGo team members even after Lee Sedol match. If DeepMind doesn't publish any paper, and retire AlphaGo immediately in March, it gonna takes more time for sure.

DeepZenGo doesn't have a perfect record but is doing pretty well on Tygem right now. Even if AlphaGo retired immediately, someone else could develop an AI that beats all humans in the next couple years.

[pookpooi]

DeepZenGo doesn't have a perfect record but is doing pretty well on Tygem right now. Even if AlphaGo retired immediately, someone else could develop an AI that beats all humans in the next couple years.

In your given situation, DeepMind also didn't published their paper, right?

[xiayun]

I got intrigued by the fast improvement of bots.

In chess, it took 10 years to improve bots who won against top humans to bots who never lose to humans. In Go, this took less than a year (or about a year when you wait for the announced games with longer time settings.) Even if you add another year because somebody might eek out a win during 2017, the time span is remarkably shorter than in chess.

I'd argue the time span isn't really remarkably shorter given how much later advanced AI got developed for Go and caught up to human. If anything, the time span for Go was remarkably longer. Technology has advanced so much, both hardware and software, since Deep Blue time. If nobody had bothered with Chess AI and the Deepmind team took it on as a challenge say in 2014, it'd probably take them way less time comparing to Go. And if they hadn't published the Nature paper, who knows how long it'd have taken existing or new AIs to get to their current levels.

[Gomoto]

In conclusion, computer go is nowhere close to optimal play.

I do not agree.

[pookpooi]

Some comment from the creator of Zen

"Yes, it is AlphaGo. I am relieved that DeepMind clarified this.

Honestly I got a little frustrated that many people didn't think that
was AlphaGo. It was almost clear to me because I know the difficulty of
developing AlphaGo-like bots.

I hope Aja can comment here, also about GodMoves

Yamato"

My comment is that DeepZenGo is struggle to improve right now, it seems to stagnant sometimes, if you look into Computer Go server, some of the newer versions are even worse than older version. So keeping progression graph linear is a very demanding task.

[Tumtumtum]

In conclusion, computer go is nowhere close to optimal play.

I do not agree.

How could you not agree? Chess softwares are 400-800 elos away from perfect play. Go is a much more complicated game.
The reason why humans could last so long in chess is just due to draws being an option. Even now one might get one.

[pookpooi]

How could you not agree? Chess softwares are 400-800 elos away from perfect play. Go is a much more complicated game.The reason why humans could last so long in chess is just due to draws being an option. Even now one might get one.

Wouldn't perfect play will always win against imperfect opponent? At least if perfect play has first-move advantage, it'll always win, so elo difference is near infinity. But if you're talking about odd match (handicap game), elo is not design for that, we just adapt and estimate (just like one stone = 230 elo in go, which is still controversial).

[Tumtumtum]

How could you not agree? Chess softwares are 400-800 elos away from perfect play. Go is a much more complicated game.The reason why humans could last so long in chess is just due to draws being an option. Even now one might get one.

Wouldn't perfect play will always win against imperfect opponent? At least if perfect play has first-move advantage, it'll always win, so elo difference is near infinity. But if you're talking about odd match (handicap game), elo is not design for that, we just adapt and estimate (just like one stone = 230 elo in go, which is still controversial).

I don't understand the point of you quoting me. Anyway, in go as long as neither is playing perfectly the winning % will most likely not be 100. In chess one needs to be very far away from perfect play to not get even one draw in a thousand games against perfect play.

[pookpooi]

I don't understand the point of you quoting me. Anyway, in go as long as neither is playing perfectly the winning % will most likely not be 100. In chess one needs to be very far away from perfect play to not get even one draw in a thousand games against perfect play.

My fault that I'm not explain clearly.
To be direct, do you have any source that conclude the Elo differece between chess AI and perfect play is between 400-800?
Maybe there's a method that can estimated that number even when the game itself is not solved yet.

[HermanHiddema]

Wouldn't perfect play will always win against imperfect opponent?

No. If I develop an AI player that misses the best move once every 1000 moves on average, then it would lose sometimes, but draw most games, provided optimal integer komi is used. If fractional komi is used, the imperfect player would even win some games. Regardless, it would score some percentage (say 45 points every 100 games) and thus it is possible to establish a stable Elo rating difference. Similarly, if other AI players miss perfect play every 500, 200, 100 or 50 moves, they would get slightly worse ratings.

[Gomoto]

The argument was the level is nowhere close to optimal play.

In fact the top players and top neural networks are quite close to optimal play.

[pookpooi]

Wouldn't perfect play will always win against imperfect opponent?

No. If I develop an AI player that misses the best move once every 1000 moves on average, then it would lose sometimes, but draw most games, provided optimal integer komi is used. If fractional komi is used, the imperfect player would even win some games. Regardless, it would score some percentage (say 45 points every 100 games) and thus it is possible to establish a stable Elo rating difference. Similarly, if other AI players miss perfect play every 500, 200, 100 or 50 moves, they would get slightly worse ratings.

Yes, you're completely right. I really messed up 'imperfect opponent' and 'imperfect play' a big time.

[Jhyn]

The argument was the level is nowhere close to optimal play.

In fact the top players and top neural networks are quite close to optimal play.

Do you have any arguments going in this direction? Without any arguments, I would find the opposite much more believable. Optimal play is a fantastic beast.