A conversation between Vlad Firoiu, an AI grad student at MIT, and Brian Hamrick. Vlad believes Alphago will win, Brian disagrees.
Tim Vlad, how much do you know about computer go AI?
In particular, i want to place online bets on Lee Sedol vs Alphago.
Vlad Firoiu I know a fair bit about it
people in my lab at MIT are also placing bets
I think alphago will win
Brian Hamrick I think I'd take lee sedol's side at 1:1
maybe like 70%
Tim so Vlad Firoiu is betting on Alphago
Brian the pro's opinions is that alphago has weaknesses in macro play
I think the weakness is that alphago is probably trained on most pro games, as opposed to mainly with top pro games
so it will have some weaknesses that come from lower tier pros
there straight up aren't that many top pro games
Tim so the claim is that the learning algorithms that deepmind uses doesn't have enough data in top pro games to train on?
Brian I think so
so there's also potential weaknesses in how alphago learned
because there's a lot of relatively set sequences of moves, and some of those aren't optimal depending on the rest of the board
but those situations almost never arise, so it's likely that alphago has never seen the reasons for the patterns
so potentially it will lose out by not deviating, or by sedol deviating
Vlad: ok, so basically it is learning to be a good but not best human
Vlad: so 1. I think that even if that were the case it could still win
by looking further ahead
which is something humans are terrible at
but computers are amazing at due to memory, parallelization, and modern tree search
Vlad: 2. alphago will bootstrap by playing itself
having reached pro level play, it can generate infinite high quality data
this has been used in other games to exceed human performance
I wouldn't be surprised if alphago reaches a level way past lee sedol or that any human will reach
Vlad: again, all these things can leverage massive parallelization and I suspect google will go all-out on this in terms of computing resources and engineering
not to mention having go experts and AI experts improving the algorithms
Vlad: 3. I think that alphago pretty much has the fundamentally "right" algorithm
this is obviously not something I can prove
but I think that the combination of heuristic learners with tree searchers is pretty darn good
Brian: hmm I don't think that bootstrapping really works the way vlad says it does
it seems like it should just make alphago more confident in something arbitrary, which may or may not be correct
depending on the specifics it may or may not be true that alphago can read further than sedol
it's probably comparable at this point
Tim read further?
Brian read = look ahead
Tim hm, how could people possibly look farther ahead than computer?
Brian well so traditionally in go it's been because humans only actually consider ~5 moves max, whereas computers don't know which 5 moves those are so they have to search a lot more
the neural net part of alphago is probably attempting to mimic that intuition of humans
so alphago may not have the same limitations as older go ai
Vlad: well brian's right about the alphago being better than older ai's by having more human-like intuition
also there are good ways to leverage these heuristics to guide the tree search
Vlad: it is true that bootstrapping isn't quite as straightforward as it seems
but deepmind is well aware of this and has several ideas on how to make it work
which they've demonstrated on simpler games
Vlad: re lookahead: brian is trying to account for lee sedol's better heuristics which allow him to look at a much smaller branching tree
which is certainly valid in comparing to older go bots which had bad heuristics
but I suspect that alphago's heuristics are already good enough to allow super-human tree search to kick in
Brian: hm I should mention that even the top go players make multiple mistakes per game (as determined by post-game analysis)
so this means two things
one is that alphago may have inherited some of those mistakes from the learning set
the other is that alphago may be able to avoid that because of the consistency of the tree-search part of the algorithm
it's very difficult to tell which side is advantaged from that
Brian: one of the reasons that it's unclear is that most of the time the reason that a move is a mistake isn't revealed until many many moves later, and there's a lot of variations in the meanwhile
so it's not clear that the tree search will be enough to recognize it
Vlad Firoiu works on deep learning at MIT, in the lab of world famous AI researcher Josh Tenenbaum. The opinions he expresses are entirely his own, and are not affiliated with MIT or Professor Tenenbaum.
Brian Hamrick is a graduate from MIT, IOI gold medalist, and all around very intelligent person. He blogs about a wide range of topics including Haskell, Go, programming, mathematics, Pokemon speedrunning. Check out his blog at Brianhamrick.com/blog
Postscript: chat with Jacob Steinhardt
Tim: do you think Alphago will beat Lee Sedol?
Jacob: i doubt it
Lee Sedol is way better than the last guy
i think it would have had to improve more in the last 6 months
than the improvement that it was relative to the previous state of the art
to be better than Lee Sedol
Tim: is this informed by your knowledge of go, by your knowledge of AI, or both?
i'm just looking at the elo ratings
Reposted from https://docs.google.com/document/d/1sdBnDVenK7YwuQGZpKsbb0pxpB-ghl4uhn0-NbBHsw0/edit?usp=sharing