AlphaGo Is Not the Solution to AI
cacm.acm.org
cacm.acm.org
By the way, if Monte Carlo search worked, then we would have seen MC beating 9-dan pro's long ago, which clearly didn't happen, so the author's definition of "effective" lacks pragmatic insight - the real limiting factor in this sort of evaluative judgment.
The author doesn't seem to have watched any of the matches of AlphaGo vs Lee Sedol, though, as much of the pre-game and post-game discussion brings to light some important details surrounding specific implementation features in AlphaGo by Deep Mind. Glossing over that makes this article remarkably uncritical and uninformative.
Edit: it's one thing to combat hype, but representing the facts as they are in contradistinction to that hype is another matter entirely. Anti-hype is perhaps more pernicious in bringing about an "AI Winter" than is drumming up interest in what AI, or more accurately AGI, will be capable of in the not-too-distant future. And given current developments, I think the only AI Winter we have to fear is the one where uninformed persons accost us with misguided fears about AI and what its capabilities will be.
It matters because it could lead to another AI winter when the hype falls through.
Once the difficulty of those problems started to be understood, of course funding dried up. Industry is focused on short-term ROI, so it's hard to get funding if the profits won't be seen for 50 years.
The difference is that now there's an entire string of profitable markets for solving near-term AI problems. AI is fundamental to the business models of some of the world's largest companies (i.e. Google). There's basically zero risk of an AI winter when we're on the verge of advanced robotics, Watson-style Q&A systems, self-driving cars, large-scale genomics, etc.
An AGI winter is another story, but most AI work isn't really focused on serious, full-scale AGI right now anyway. That winter never ended and is ongoing. Everyone is focused on incremental AI improvements because no one really knows what's required for full AGI, and are in the meantime hoping they'll hit on it while building on known techniques like deep learning, comp neurosci, etc.
tl;dr: As long as investors continue to see marketable products for new AI developments anywhere in the next ~10 years, funding won't dry up again.
As someone who experienced AI winter (and ended up leaving the field) and even attended the AAAI panel where the term was coined on stage, I am not so confident.
What happened then looks like it could happen again. At that time there were tons of expert systems being deployed with high promises that they would revolutionize business and practical problem solving. Non-technical people from other fields were flooding in to sell "solutions." Dedicated workstations were proliferating.
The self-driving car hype is already at the same level as the expert systems hype was back then.
> tl;dr: As long as investors continue to see marketable products for new AI developments anywhere in the next ~10 years, funding won't dry up again.
Well sure, that's true, but no different from saying "as long as house prices continue to go up there will be a housing boom."
Reminds me of this XKCD: http://xkcd.com/1425/
It is happening again, "Dear Aunt, Let's Set So Double The Killer Delete Select All", Honda selling lane following and adaptive cruise control labeled as self driving (so does tesla btw), etc. There are plenty of scam products and misleading marketing surrounding AI right now.
I agree with the main point about Monte Carlo methods here.
Neural networks are a "dumb" way to stack another layer of optimisation on top of many algorithms.
I won't lie there is a certain bitterness watching this having used the same principles on less gimmicky problems which I'm sure the author shares
AI is massively popular in both industry and academia at the moment. It is at no risk of being deinflated by "anti-hype".
Isn't it precisely when something is massively inflated that it's most at risk of being deflated?
Remember those mid-80s films with intelligent computers everywhere? People really thought that AI would be a solved problem within years. Then reality struck and the AI winter fell.
The challenge is that to approximately 99.99% of the seven billion people on this planet, including roughly every journalist, the recent progress made by AI researchers is indistinguishable from "magic."
When all those people try to reason about "magical things," they're bound to reach nonsensical conclusions and say nonsensical things. To them, debating on the promise of technologies like deep learning is a lot like debating how many angels can dance on the head of a pin.[1]
--
[1] http://www.straightdope.com/columns/read/1008/did-medieval-s...
And while risking another AI Winter is no fun, it's possible that AI researchers are probably going to be sorely tempted to participate in the hype cycle as well. In fact, considering how much money is being poured into AI research, what is it to say that the hype has already just begun?
Games of incomplete information, like poker, are much harder, if not impossible for computers to master without actual AI. It was only recently that heads-up, limit poker was "solved". Even in this case, the "solution" meant only that the computer could put forth a strategy that would not lose significantly over time, not a winning strategy. Games like no-limit hold em are infinitely more complex. This is true of heads-up, but vastly more so for full-ring games.
AlphaGo's victory has certainly been very impressive, but its also predictable. Given enough computing power, winning any game of complete information is just a matter of brute computing force. We'll see how computers (and real AI) can do in the future in regards to other games.
That's like saying public-key encryption is broken because we can brute force decryption.
When we talk about solving a game (or solving an optimization problem, etc), we aren't simply interested in whether the problem is computable. We want a efficient, polynomial-time algorithm if we're interested in the generalized problem (the size of the board is a parameter), otherwise at least something that runs and completes in a reasonable time if we're OK with solving only a particular board size (an instance, in terms of computational complexity).
The generalized go problem is PSPACE-hard or EXPTIME-complete depending on the ruleset you adopt. So it's not just a matter of brute processing power.
Now, most people outside theoretical computer science research don't care much about generalized games and asymptotics, they're fine with solving 19x19 Go instead of n x n Go. But that doesn't help you much: the only board size that has been solved is 5x5 [0], and there are only about 10^10 legal positions in a 5x5 board. The standard 19x19 has 10^170 legal positions! Good luck brute forcing it.
No, that's like saying that public-key encryption will eventually be broken because at the end of the day its just about having enough brute force processing.
>When we talk about solving a game (or solving an optimization problem, etc), we aren't simply interested in whether the problem is computable.
No we aren't. But games of complete information that ARE computable, such as chess and GO can be removed from discussion. There is no question that these games of complete information are "solvable" with or without AI, since every move, from start to finish, can be calculated before the game even begins - its just a matter of processing power.
Which brings us to the question of AI and games of incomplete information. By definition full-ring no-limit hold-em is not computable. Even with infinite processing power, with every possible permutation calculated, there is no guarantee that a brute-force computer player is going to win full-ring no-limit hold-em. If there are multiple unknown players to act behind you, there is no way to compute with certainty what the outcome will be. This is where AI becomes really interesting as far as gaming is concerned.
In a theoretical universe, sure. In the one that we live in - no.
Let's assume that we can turn every atom in the entire universe that we know of into a 4Ghz processor, further assume that we can perfectly parallelize the search and that we can check one board position in one cycle on one of those processors.
How long would it take to check all possible 19x19 go positions?
(1 * 10^80 processors) * (4 * 10^9 checks) = 10^89 checks per second.
Total number of valid go board states: 10^170, therefore it would still take 10^79 seconds to check all the positions in order to find a perfect game.
Clearly, it is not possible to ever brute-force go, no matter what kind of advances computing technology will take.
Many times before, claims like this have been disproved.
It's even true for rock-paper-scissors and the "solution" is trivial, and there is no better strategy. The lack of winning strategy can be the property of the game itself. It's entirely not trivial if a "winning strategy" to a game like poker even exists.
If by "brute force" you mean exploring all possible moves then in theory yes, in practice the computing power and time is larger than the universe for chess or go, so no.
There is no significance to AlphaGo - to people paying attention to AI research. The methods aren't incredibly novel, the idea that CNNs could make very good value approximators isn't new, etc. I predicted computers would be beating expert Go players by the end of 2015, and was actually really disappointed when I was wrong. Turns out I wasn't, but it had just been kept secret.
However AlphaGo's victory is significant to the general public and the skeptics. It's a sign of the rapid progress being made in AI over the last few years. Rapid progress that shows no sign of halting.
Some of the people I spoke to about my Go prediction thought that it might be possible, but was probably too hard a domain. Hopefully those people will update their beliefs and be a little more optimistic about NNs, and AI in general. They are extremely powerful tools that we are just figuring out how to use. Go will hardly be the last domain they succeed at.
And who are you? Predictions are a dime a dozen; what gives any relevance to yours?
Predict weather 2 weeks out to within 1 degree at 50/50 accuracy.
Do face recognition well enough to identify people from surveillance videos. (from the set off all known criminals with 50/50 accuracy including false positives and false negatives.)
Program a computer well enough to replace a Jr developer.
Write a book good enough to become popular on its own merits.
Replace in person translators for real time language X to Y langue translation at the UN.
There is going to be no "solution" to AI.
Only emerging complexity which I am guessing at one point is going to give us something akin a semi aware entity. After that all bets are off.
I actually think it's OK if some people think AlphaGo is more general purpose than it really is, because the general attitude or at least appropriate response to the likely development of AGI relatively soon still seems to be amazingly muted.
The potential for neural networks to learn to play games, by playing each other, has been apparent at least since NeuroGammon (https://en.wikipedia.org/wiki/Neurogammon), in 1989. The "playing each other" mechanism is key, because it allows large-scale training.
The question is, what games happen to be well structured to allow so-called "credit assignment" so that you can know what states lead to wins across a reasonable spectrum of (unknown/unknowable) opponent plays.
According to the OP, it is known that you can do credit assignment in Go. ("For Go itself, it has been well-known for a decade that Monte Carlo tree search (i.e., valuation by assuming randomized playout) is unusually effective in Go.") I did not know that, but if so, it would simplify the Go problem, although the search space remains huge, etc.
The key advancement in AlphaGo seems to be generating a subset good moves, as opposed to randomly selecting them, and determining whether a given configuration has a probability of winning. Monte Carlo Tree Search by itself only gets you to amateur ranks from what I read, but won't beat a pro player. Being able to combine the policy and value networks, and figuring out how to bootstrap and train them was where all the work is.
In hindsight, everything looks like an obvious combination of separate already-known-to-be-good pieces. I don't think this achievement should be hand waved away.
More importantly, the DeepMind folks are being pretty humble about it and not making the big claims everything seems to think they're making.
https://en.wikipedia.org/wiki/Universal_approximation_theore...
Neural networks are an algorithmically simple way to approximate any mathematical function. So, really, why _wouldn't_ you think this?
Ultimately, almost all boardgames can be reduced to mathematical functions.
Awareness is a virtue in and of itself, and any machine that has it will require an implementation of awareness itself, and that is regardless of the machine's intelligence. Awareness is about self motivation and the creating of its own goals. It's the generating of meaning to back one's words as well as actions, as even we often struggle to do.
As far as I am aware, no one at Deep Mind or IBM is actively working on intent and self-awareness. So for the foreseeable future, if a two-legged drone shows up at your door and points a gun at you, it won't be because AI has awakened. It'll be because someone at the CIA wants to kill you.
Other discussion from Yann Lecun https://www.facebook.com/yann.lecun/posts/10153426027052143 and http://www.information-age.com/technology/applications-and-d...
"Tesler's Theorem (ca. 1970). My formulation of what others have since called the “AI Effect”. As commonly quoted: “Artificial Intelligence is whatever hasn't been done yet”. What I actually said was: “Intelligence is whatever machines haven't done yet”."