AlphaGo vs Lee Sedol, who will win?

[Uberdude]

In March AlphaGo, the new neural-net based AI from Google Deepmind which has already beaten Fan Hui 2p 5-0, will play Lee Sedol 9p, who probably has the best claim to be the top pro in the world over the last decade. We still don't know details (should be 5 games) like time limits, but if we assume at least the 1 hour main time from the Fan Hui match (I hope at least 3 hours) who do you think will win?

https://googleblog.blogspot.co.uk/2016/ ... me-go.html

(discussion thread about AlphaGo here: viewtopic.php?f=18&p=198226#p198226)

[hyperpape]

My vote depends on how much they can improve their machine since October. I don't think the 5-0 victory against Fan Hui or what they've said about beating other bots (80% against Crazy Stone, Zen and Pachi with 4 stones) is enough to make us think they'll beat Lee. But with five months to train their bot, they may still close the gap.

But based on what we've seen so far, I voted 4-1 for Lee.

[Uberdude]

hyperpape, that's a good point. If it was around 3d back in February 2015 when Charles went to that talk, and let's say Euro 8d (or more?) when it played Fan Hui in October 2015 (8 months later), who knows how much stronger it will be by March 2016 (5 months later). I expect diminishing returns, but that is still mightily impressive improvement. Anyone spotted some mysterious new Tygem 9ds with an AlphaGo style (whatever that is)?

[sparky314]

I'd agree with hyperpape, it really depends on how much they improve the engine before March. I voted too close to call. I think Lee Sedol has an advantage from the information we have about the engine back in October, but with 5 additional months of development and training? There's no way to tell how it's going to go.

[Solomon]

I am expecting AlphaGo to have reached two more super saiyan levels since October, and will win against Lee Sedol 5-0 just like it did against Fan Hui. It will be majestic.

[jeromie]

The fact that the majority of voters expect Lee Sedol to lose at least one even game to a computer opponent is a telling sign of just how far Go software has advanced and the efficacy of the press coverage surrounding the loss to Fan Hui.

[Elom]

Humans can make mistakes. It feels very realistic that AlphaGo will improve significantly (albeit a much slower rate than previously) and surely win one game or maybe 2, mainly due to the fact that it's the first time playing AlphaGo, but even that I doubt from an experienced world champion. But it's still unclear whether the "barrier" to top level play can be broken. If this was to be held at sometime in late 2016, most normal people would have little feasible idea of how things would turn out, but now it seems to depend a lot on the stability of a experienced human, playing against a new program that is not designed to work best out-of-the-box, being more like a yunzi-stone.

That said, while Lee 9p may be able to analyse games from Alpha-go, I doubt it would have as much effect as studying other classic programs, while alpha-go will be constantly analysing 100's of games. But I can still only imagine Alphago possibly winning one game in it's current state, which is still a huge achievement AI Go.

A pro can be of a level to crush a weaker pro and then be crushed by a top pro easily. There's a lot of probability involved in assessing AlphaGo's true level.

[xed_over]

why can't we have a discussion without hiding them?

[hyperpape]

why can't we have a discussion without hiding them?

I didn't want to spoil the poll, I guess? Seemed like an ok thing to do.

[emeraldemon]

I mentioned in the other thread that goratings.org thinks Fan Hui would have a 3% chance to win against Lee Sedol.

AlphaGo seems better than Fan Hui, but we don't know by how much. Let's be conservative and say it has a 5% chance of beating Lee Sedol.

The odds that Lee wins:
5/5: 77%
4/5: 20%
3/5: 2%
less: <1%

Goratings only has 11 games from Fan Hui (2 wins 9 losses) and the last game was December 2014. The rating algorithm probably has considerable uncertainty about his true rating. Also the 5-0 isn't a particularly large sample size either. It's still possible the odds for AlphaGo could be <1%. And of course we could also argue the program will continue to improve and the odds are much higher than 5%.

[Joaz Banbeck]

At his best, Lee will crush any computer. So I am tempted to predict 5-0 in his favor.

But I'll give the computer one win when Lee makes a mistake.

My vote is Lee wins 4-1.

[wineandgolover]

I mentioned in the other thread that goratings.org thinks Fan Hui would have a 3% chance to win against Lee Sedol.

AlphaGo seems better than Fan Hui, but we don't know by how much. Let's be conservative and say it has a 5% chance of beating Lee Sedol.

I would redo your math. How much stronger than Fan Hui is AlphaGo likely to be? Asked another way, what would it's likelihood to win a single game be, if it's odds of wInning five in a row are 50%? Say, x^5=.5

I think that comes out to an 87% likelihood.

Now for you rating experts, what does that mean in elo terms versus Fan Hui? And then for its chances against Lee Sedol?

That's if the odds of a sweep were, indeed, 50-50. It could be far stronger, increasing the odds of the same result. So this is more of a floor for estimating AlphaGo's strength.

(Sorry, I'm literally on a desert island, and this is all I had time for.)

[uPWarrior]

I would redo your math. How much stronger than Fan Hui is AlphaGo likely to be? Asked another way, what would it's likelihood to win a single game be, if it's odds of wInning five in a row are 50%? Say, x^5=.5

This question might be partially answered in the paper itself.

According to this interesting figure:


The distributed version of AlphaGo is ~250 rating points above Fan Hui. If my calculations are right, this would imply a 96.3% probability of winning a single game and an 82.6% probability of winning 5 games in a row.

(It also shows that pro-level strength in our home computers might be closer than we think)

[Calvin Clark]

There are various reasons I think AlphaGo will win, despite the fact that I've started to read comments by pros suggesting otherwise.

1. I don't think the funders of the Google Deepmind project have any intention of giving Lee a million dollars, no matter how much they'd like to be VIPs at the bar he will certainly open when he retires (again).

2.It's really easy to test a program's improvements or hardware improvements in secret on the Internet. If we see no more AlphaGo games before the match in March, then there is precious little to go on for Lee, whereas everyone has access to a lot of Lee's games.

3.Lee will not get stronger between now and March, but there is every reason to believe that AlpahaGo will.

[Firebrand]

Various pros favor Lee Sedol at the moment, so they clearly see significant difference between his level and what AlphaGo showed in its match against Fan Hui. The only unknown quantity is AlphaGo's improvement between October and March. It brings us to question of how long DeepMind team waited and trained AlphaGo before offering Fan to play against it.
If they had no time limit at all, it's likely that current version of the program learned about as much as it can (because obviously these guys want their creation to be at its best in any public games). Of course, there will be some improvements, but they'll probably be quite small - unless the team had a sudden breakthrough.
If they had a time limit from Google (who, in example, may have wanted their Go AI to hit the news before the one made by Facebook did), then it's entirely possible for March AlphaGo to be on a completely different level from October version.
I don't know enough about people and companies involved to make a good guess which scenario is more likely, so it seems wiser not to make any guess at all. Personally, I'd prefer slow progress, so I can see more interesting games between top pros and computers before the latter completely overcome the former.