AlphaGo beats Lee Sedol again in match 2 of 5
gogameguru.com
gogameguru.com
AlphaGo plays some unusual moves that go clearly against any classically trained Go players. Moves that simply don't quite fit into the current theories of Go playing, and the world's top players are struggling to explain what's the purpose/strategy behind them.
I've been giving it some thought. When I was learning to play Go as a teenager in China, I followed a fairly standard, classical learning path. First I learned the rules, then progressively I learn the more abstract theories and tactics. Many of these theories, as I see them now, draw analogies from the physical world, and are used as tools to hide the underlying complexity (chunking), and enable the players to think at a higher level.
For example, we're taught of considering connected stones as one unit, and give this one unit attributes like dead, alive, strong, weak, projecting influence in the surrounding areas. In other words, much like a standalone army unit.
These abstractions all made a lot of sense, and feels natural, and certainly helps game play -- no player can consider the dozens (sometimes over 100) stones all as individuals and come up with a coherent game play. Chunking is such a natural and useful way of thinking.
But watching AlphaGo, I am not sure that's how it thinks of the game. Maybe it simply doesn't do chunking at all, or maybe it does chunking its own way, not influenced by the physical world as we humans invariably do. AlphaGo's moves are sometimes strange, and couldn't be explained by the way humans chunk the game.
It's both exciting and eerie. It's like another intelligent species opening up a new way of looking at the world (at least for this very specific domain). and much to our surprise, it's a new way that's more powerful than ours.
(This also applies without a 100% chance of winning, as long as its chances of winning hover near the highest percent it's able to distinguish.)
Even if it did output an actual 100% chance, AlphaGo would still end up picking moves favored by the policy network, so it would probably just revert to playing like it predicts a human pro would.
It's similar to how ray tracing renderers start to return weird speckle patterns when the room is dark enough.
And the policy network chooses branches to investigate, not which one to choose. It adds sample resolution to places pros might play, but doesn't add to the estimated probability of winning.
Edit: Actually, since places pros might play have higher sample resolution, they're less random. So worse moves get worse evaluation, and a higher chance of leading the pack. This might actually bias AlphaGo to play some pretty bad moves - but, again, this is all assuming it's going to win anyway.
Pretty much the same thing happened with TD-Gammon with it playing unconventional moves, in the longer term humans ended up adopting some of TD-Gammon's tactics once they understood how they played out, it wouldn't be surprising to see the same happen with Go.
“Top competitors who once relied on particular styles of play are now forced to mix up their strategies, for fear that powerful analysis engines will be used to reveal fatal weaknesses in favoured openings....Anything unusual that you can produce has quadruple, quintuple the value, precisely because your opponent is likely to do the predictable stuff, which is on a computer” [1]
Anand isn't really talking about strategy here, he's just talking about choice of opening. Players with narrow opening repertoires, like Fischer, have always been easier to prepare for than players who play a wide variety of openings.
As far as actual changes to strategy, the most obvious one is that computers tend to value material more highly than humans. So a computer will take a risky pawn if it looks sound, while a human will see that taking the pawn is very complicated and prefer a simpler move.
(1) Online game databases have made it easier for players to track developments in opening theory and prepare to play specific opponents
(2) Chess engines add to this be used to search for antidotes to complicated opening systems
(3) Young players have greater access to high-quality sparring partners - either engines or fellow humans on online servers.
This has lead to the best players becoming younger, and players playing more varied and less 'sharp' openings.
Citation needed.
Looks like citation 46 is the relevant one here.
It uses MCTS, which is unlike minimax. It doesn't use temporal difference learning, although they say that the policy somewhat resembles TD.
That doesn't sound like 'essentially built on', its sounds maybe like 'slightly influenced by'
Tesauro's work on TD-Gammon was pioneering at the high level, i.e. combining reinforcement learning + self-play + neural networks.
I fixed this behavior by scoring earlier wins higher than later wins. Now it will actually finish games (and win), but almost invariably its edge is very small, no matter how well or poorly I play. Because of the new win scoring, it willingly sacrifices its own advantage if it means securing a win even one turn earlier. (And since scoring is symmetrical, this has the added advantage of working to delay any win it sees for me, thus increasing the possibility of me making a mistake!)
I suppose I could try modifying the scoring rules again, to weight them by positional advantage. A "show off" mode if you like :) And again, with the flip side of working to create the least humiliating losses for itself.
Absolutely. Also worth noting that it may be simply unable to distinguish between good and bad moves if both outcomes lead to a win, since it has no conception of the margin of victory being important.
So it might not be that it increased win probability, but that both paths led to 100% win probability and it started playing "stupidly" due to lacking a score-maximizing bias.
Humans, I think, have the natural instinct to "hedge" themselves in games like go and chess, by creating positional/material advantages now to offset unknowns later. Of course, that advantage becomes useless in the end game, when all that matters is the binary win/lose.
An AI, which may have a deeper/broader view of the game tree than its human opponent (despite evaluating individual position strength in roughly the same manner), may see less of a need to "hedge" now, and instead spend moves creating more of a guaranteed advantage later (as you suggest). And indeed, my experience with my AI is that during the endgame (in which an AI generally knows with certainty the eventual outcome of each of its moves), it tends to retain the smallest advantage possible to win, preferring instead to spend moves to win sooner.
That's actually an excellent way to win chess games. Keep your eye on the mate while the other person is focusing on position and material.
I'm confused. Why would 'make the winning move' not be the way to maximise probability of winning?
I suppose that, in Hive, it is more likely that a path to a win is longer rather than shorter. Hence, when my AI was arbitrarily choosing "winning" moves, it statistically chose those that drew the game out.
Your post should be required reading in this discussion.
People forget how literal computers are.
That's called programming
if you have fully autonomous robots which can fight your war, you'd be able to launch a massive offensive within hours. properly mobilizing defenses and responding to that invasion would take too long, as any command centers would've already been wiped out by the first attack.
Humans play that way too. Everyone wants to maximize the chance of leading by >=1 stone. The difference is that AlphaGo is better at calculating a precise value of a position, so that when uncertainly plays in, AlphaGo can play for, say, "1-3 stone lead", while a human can only get confidence in "1-7 stone lead", and thus needs to play excessively aggressively to overcome the uncertainty.
For instance, I developed a system that used machine learning and linear solver models to spit out a series of actions to take in response to some events. The actions were to be acted on by humans who were experts in the field. In fact, they were the ones from whom we inferred the relevant initial heuristics.
Everyday, I would get a support call from one of the users. They'd be like, 'this output is completely wrong. You have a bug in your code.'
I'd then have to spend several hours walking through each of the actions with them and recording the results. In every case, the machine would produce recommended actions that were optimal. However, they were rarely intuitive.
In the end, it took months of this back and forth until the experts began to trust the machine outputs.
This is the frightening thing about AI - not only can an AI outperform experts, but it often makes decisions that are incomprehensible.
If I'm an expert in some domain and a computer is telling me to do something completely different ("Trust me--just drive over the river!") I'm certainly going to question the result.
> I sense a change in the announcer's attitude towards AlphaGo. Yesterday there were a few strange moves from AlphaGo that were called mistakes; today, similar moves were called "interesting".
Later, he did admit that the "overextension" on the north side of the board was more solid than he originally thought, and called it a good move.
He never explicitly said that a move was "good" or "bad", and always emphasized that as he was talking, his analysis of the game was relatively shallow compared to the players. But in hindsight, whenever he point out an "bad-juju feel" on the part of Lee's move, AlphaGo managed to find a way to attack the position.
Overall, you knew when either player made a good move, because the commentator would stop talking and just stare at the board for minutes, at least until the other commentator (an amateur player) would force a conversation, so that the feed wouldn't be quiet.
The vast, vast majority of the time, the English-speaking 9-dan was predicting the moves of both players, in positions more complicated than I could read. (Oh, but it was obvious both players would move there. There were clearly times when the commentator would veer off into a deep distant conversation with the predicted moves still on the demonstration board, because he KNEW both players were going to play out a sequence of maybe 6 or 7 moves forward).
They really got a world-class commentator on the English live feed. If you got 4 hours to spare, I suggest watching the game live.
I think this will the theme of our future interactions with AIs. We simply can't imagine in advance how they will see and interact with the world. There will be many surprises.
I have been watching Myungwan Kim's commentary for the games - and it seems notable that a few moves he finds very peculiar immediately when they are made, he will later point out to as achieving very good results some 20 moves later. So it also seems quite possible that AlphaGo is actually reading this far ahead, to find those peculiar moves achieve better results than from the more standard approaches.
Whether these constitute a 'new way' or not I think depends highly on whether these kind of moves can fit into some general heuristics useful for considering positions, or whether the ability to make them is limited to intelligence's with extremely high computational power for reading ahead.
This. It's a fairly common feature of any AI that uses some form of tree search/minimax, and the effect is very pronounced in chess. Even the best human players can only think 6-8 plies into the feature versus ~18 for a computer. What we can (could?) do is apply smarter evaluation functions to the board states resulting from candidate plays and stop considering moves that look problematic earlier in the search (game tree pruning). AI tends to use very simple evaluation functions that can be computed quickly. They do so given that 1) it allows for deeper search, and a weak heuristic evaluated far in the future often beats a strong one evaluated a few plies prior and 2) for some games (like Go) it's really hard to codify the "intuitions" that human players speak of.
Because search based AI considers board states __very__ far in the future, the results are often completely counterintuitive in a game with an established theory of play. Those theories are born of humans, for humans.
The introduction of MCTS some years back was the first leap towards a human level Go AI (incidentally, MCTS is more human-like than exhaustive tree search in that it prunes aggressively by making early judgement calls as to what merits further consideration). AlphaGo's use of deep policy and evaluation networks to score the board is very cool, and the next step in that journey. What's interesting to me is that, unlike chess AI, AlphaGo might actually advance the human theory of Go. It's possible that these "strange moves" will lead to some very interesting insights if DeepMind traces them through the eval and policy networks and manages to back out a more general theory of play.
I think that Chess machines play perfectly for the next 8 moves, but don't necessarily sense the importance of a Knight Outpost (which may have relevance 20 moves ahead. A proper Knight Outpost will remain a fork threat for the rest of the game).
It is far easier for a Human to beat a Chess Machine at positional play (ex: a backwards pawn shape will probably be a problem at endgame, 30+ moves from now) than to beat a Chess Machine at tactical play (3 moves from now, I can force a fork between two minor pieces)
Do some reading on Stockfish for example if you doubt the veracity of my statement.
But its just as you say: its weighting parameters and heuristics. When Stockfish recognizes a backwards pawn, it deducts a point value. When Stockfish recognizes "pawn on 6th row", it adds a point value to that pawn.
But that's a heuristic. A trained heuristic using games, but still comes down to what I understand to be a +/- point value (like... +35 centipawns).
In contrast, a chess engine truly knows that if you do X move, it will force a Rook / Minor piece exchange in 8 moves.
When you play positionally vs Stockfish, you're arguing with a heuristic (a heuristic which has been refined over many cycles of machine learning, but a heuristic nonetheless that comes down to "+/- centipawns") . When you play tactically vs Stockfish, it is evaluating positions more than a dozen moves ahead of what is humanly possible.
When you play against Stockfish in endgame tablebase mode, it plays utterly, and provably, perfectly.
Take a pick of what game you want to play against it. IMO, I'd bet on its positional "weakness" (yes, it is still very strong at positional play, but it is the most "heuristical" part of the engine)
My experience is with Chess and Chess AI, but in my experience, the more positional knowledge built into the evaluation function, the better the search performs, even if you have to sacrifice some speed for more thorough evaluation. A significant positional weakness may never be discovered within the search horizon of a chess engine because it may take 50 moves for the weakness to create a material loss, so while it's certainly possible that a deep, but carefully pruned search is being utilized, I suspect that some of the Value Network's evaluation is helping to create some of these seemingly odd moves.
For AlphaGo to recognize a position that doesn't achieve a good result for 20 moves, it would often have to search much deeper than those 20 moves (I'm not sure if you're using the term moves to mean ply or both players moving, but if it takes 20 AlphaGo moves for the advantage to materialize, that would be a minimum 40 ply search) to quiesce the search to the point that material exchanges have stopped (again, this is how chess typically does it, I don't know about Go), so the evaluation at the end of the 20 move sequence is arguably more important than a deep search. The sooner you can recognize that a position is good or bad for you, the more time you have to improve the position.
Could AlphaGO be winning in a way similar to left handed fencers having an advantage over right handers by wrong footing them rather than simply being better? Would giving Lee more chance to see this style give him a chance to catch up?
Think Bruce Lee and the creation of Jeet Kune Do. Before him everyone concentrated on improving one style by following it classically, rather than just thinking of 'how do I defeat someone'.
IMHO Lee is the best at the current style of Go. AlphaGO is the best at playing Go. Maybe humans can devise a better style and defeat AlphaGo, but I'm sure AlphaGo can adapt easily if another style exists.
Swimming. It used to be that swimmers were supposed to be streamlined and avoid bulky muscles. Then a weightlifter decided he wanted to swim. Swimmers today all lift weights.
Programming. It used to be that people built programs in a very top down, heavily planned way. Think waterfall. We now understand that a highly iterative process is more appropriate in most areas of programming.
Expert systems. It used to be that we would develop expert systems (machine translation, competitive games, etc) through building large sets of explicit rules based on what human experts thought would work. Today we start with simple systems, large data sets, and use a variety of machine learning algorithms to let the program figure out its own rules. (One of the giant turning points there was when Google Translate completely demolished all existing translation software.)
http://www.nytimes.com/2006/08/20/sports/playmagazine/20fede...
Although it takes a few paragraphs until it gets into the details of "today's power-baseline game."
Nowadays, top players slug it out baseline-to-baseline.
In terms of stance, we were taught to hit from a rotated position where your shoulder faces the net, and a normal vector from your chest points to either the left or right side of the court.
Nowadays, it's much more common to hit from an "open" position, where your body is facing the net, not turned. This would have been considered "unprepared" or poor footwork in my day, but it actually allows for greater reach. It does make it more difficult to hit a hard shot, but that's made up for by racquet technology and generally stronger players.
Which is a curious point. The gripes about early brute force search algorithms (e.g. Deep Blue?) were that they felt unnature.
However, as the searches get more nuanced and finely grained, is there a point at which a fast machine begins doing fast stupid machine things quickly enough to feel smart?
Are there any chess / Go analogs of the Turing test? Or is a computer players always still recognizable at a high level?
A Turing test for game players is an interesting idea, it would be useful for designing game players that are good sparring partners rather than brutes that can whipe the floor with you.
Ke Jie is an arrogant 18 year old and he's been saying on social network in the past couple days how he will defeat AlphaGo.
As for JKD, people are drawn in by its oriental esotericism, but there's no evidence it is an especially effective fighting style, or that it has something that (kick)boxing does not.
I could easily see the difference in tournaments with other clubs that were not used to left handed players.
Remember that AlphaGo has spent months developing its own style and theory of the game in a way that no human has ever seen. Its style is sure to have weaknesses, but humans will have a hard time figuring them out on first sight.
Similarly chess computers do better in some positions than others (they love open tactics!) and one of the games that Kasparov won against Deep Blue he won by playing an extreme anti-silicon style that took advantage of computer weaknesses. However Kasparov didn't have to figure out what that style was because there was a lot of knowledge floating around about how to do that.
Therefore I'd expect that Lee Sedol from a year from now could beat AlphaGo from today. And human Go will improve in general from trying to figure out what AlphaGo has discovered.
However that won't help humans going forward. AlphaGo is not done figuring out the game. At its current rate of improvement, AlphaGo a year from now, running on a single PC, should be able to beat the full distributed version of AlphaGo that is playing today. Now the march of progress is not whether computers can beat professionals. It is going to be how small a computing device can be and still beat the best player in the world.
Weaknesses are only relative to capabilities of the opponent to exploit them. If a tank has a weak spot that rockets can hit, but it's being opposed by humans on horseback, is it really a weakness in that context?
Additionally AlphaGo has the advantage that it started with a database of human play, so it has some ideas what kinds of positions humans miscalculate.
As for your tank vs horseback analogy, that's flawed at the moment. AlphaGo is probably reasonably close in strength to the human facing him. Improved human knowledge could tip the balance.
However in the future it will become an apt analogy. Computers are going to become so good that knowing the relative weaknesses in their style of play may reduce the handicap you need against them, but won't give you a chance of becoming even with them. That happened close to 20 years ago in chess, and is now only a question of time in Go.
Yes. A representation of ladders is among the input features of its neural networks.
https://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf
Stone colour 3 Player stone / opponent stone / empty
Ones 1 A constant plane filled with 1
Turns since 8 How many turns since a move was played
Liberties 8 Number of liberties (empty adjacent points)
Capture size 8 How many opponent stones would be captured
Self-atari size 8 How many of own stones would be captured
Liberties after move 8 Number of liberties after this move is played
Ladder capture 1 Whether a move at this point is a successful ladder capture
Ladder escape 1 Whether a move at this point is a successful ladder escape
Sensibleness 1 Whether a move is legal and does not fill its own eyes
Zeros 1 A constant plane filled with 0
Player color 1 Whether current player is black
(The number is how many 19x19 planes the feature consists of.)
Until someone got better weapons and suddenly the "rules" of the battlefield that dictated standing in lines across each other made no sense to follow anymore because the original principles that dictated those rules to be good were not valid anymore.
And this is just the beginning with AlphaGo. As we keep on training Deep Learning systems for other domains, we'll realise how differently they approach problems and solve them. It'll, in turn, help us in adapting these different perspectives and applying them to solve other problems as well.
It's not like this at all; let's not do this sort of thing. Humans are inveterate myth makers (viz. your description of how people conceive the Go board as army units), and our impositions on the world are easily confused for reality.
In this case, there's no "intelligent species" at work other than humans. We made this, and it is not an intelligence, it is a series of mathematical optimization functions. We have been doing this for decades, and these systems, while sophisticated, are mathematical toys that we have applied. We built and trained this thing to do exactly this.
As a student of AI you know that convolutional neural networks are black boxes and are hard to interpret. A different choice of machine would have yielded more insight about how it is operating (for example, decision trees are easier to interpret). The inscrutability of the system is not a product of its complexity; even a simple neural network is hard to understand.
This, actually, is my primary objection to using CNNs as the basic unit of machine learning - they don't help US learn, they require us to put our faith in machines that are trained to operate in ways that are resistant to inspection. In the future I hope that this research will move more towards models that provide interpretable results, so they ARE actually a tool for improved understanding.
I think the answer must be in figuring out how to decompose the black box of a CNN - it is, after all, just a set of simple algebraic operations at work, and we should be able to get something out of inspection.
I have to imagine Hinton et al. have done work in this regard, but this is far afield for me, so if it exists I don't know it.
You can say the same about your mind too which is a bunch of optimization nodes. If something is intelligent, does it matter if it's evolved in nature or created by a species who is evolved in nature?
> In the future I hope that this research will move more towards models that provide interpretable results I think it's not really possible to understand in detail how these networks operate on the level of nodes, because emergent behavior is necessarily more complex than the sum of its parts.
A CNN is a pure mathematical function - if you want, you could write it down that way. Given a set of inputs, it will always produce the same output. We don't call a linear regression model an "intelligence", a CNN is no different.
Of course I agree that humans are built up of billions of tiny machines like this, but let's appreciate the vast difference in scale.
> A CNN is a pure mathematical function That's their basic property, but who are we to say that our cell based neural network is superior? Cells are just compositions of atoms and they are defined by quantum mechanics, which is... "just" math and information.
I also think that Go might be a great communication tool between AI and humans. If you look at the commentary from this angle if's fun to think about like this.
Human intuition and to certain extent, creativity are like this as well.
"oh what if the machine suddenly came alive!?" has been done 1000 times. But such concepts like: a computer can detect and act patterns which we cannot, in ways that are almost, if not possibly intelligence, are magnitudes more believable, and therefore, compelling.
Thanks! :-)
Of course, those fools underestimated it. They should have known better...
Some examples of 5th line early shoulder hits in recent professional play - these situations are not the same as the one seen in today's game, but something like a 5th line shoulder hit is always going to be highly contextual and creative.
http://ps.waltheri.net/database/game/26929/ (move 23) http://ps.waltheri.net/database/game/69545/ (move 22) http://ps.waltheri.net/database/game/71408/ (move 22) http://ps.waltheri.net/database/game/4663/ (move 9)
The only one I can't parse is the last one. There are a lot of variations where I want to know what black's plan is.
The excellent point you're making applies in general to nearly every type of human thinking.
The way we think about other people, our intuitions about probabilities, our predictions about politics, and so on -- all are based on our peculiarly effective, yet woefully approximate, analogy based reasoning.
It shouldn't be surprising in the least when commonly accepted "expert" heuristics are proved wrong by AIs that actually search the space of possibilities with orders of magnitude more depth than we can. What's surprising -- and I think still a mystery -- is how human heuristics are able to perform so well to begin with.
I'm not a Go player, but I saw this same phenomenon as poker bots have surpassed humans in ability. As with AlphaGo, they make plays that fly in the face of years of "expert" wisdom. Of course, as with any revolutionary thinking, some of the new strategies are "obvious" in hindsight, and experts now use them. Others seem to require the computational precision of a computer to be effective in practice, and so can't be co-opted. That is, we can't extract a new human-compatible "heuristic" from them -- the complexity is just irreducible.
They are peculiarly effective only because of lack of comparison. Humans have been the most intelligent species on this planet for millennia, where no other species come even close. We don't know how ineffective those strategies are seen by a more advanced species. Well, until now.
Of course, the counterpoint could be that it's only the case because humans, with their laughable reasoning abilities, are the ones programming those computers.
AlphaGo can’t decide that it’s bored and go skydiving. Humans aren’t merely capable of playing Go. And when they do it, they can also pace around the table, and drink something, all at the same time, on a ridiculously low energy budget. Or they can decide never to learn Go in the first place but to master an equally difficult other discipline. They continuously decide what out of all of this to do at any given moment.
AlphaGo was built by humans, for a purpose selected by humans, out of algorithms designed by humans. It is not a more advanced species. It’s not even a general intelligence.
Your own original point was much better than the one made in response.
http://www.nybooks.com/articles/2010/02/11/the-chess-master-...
Which attempts to visualize machine areas of attention that look like: http://www.wildml.com/wp-content/uploads/2015/12/Screen-Shot...
.. that we'll be probably unable to comprehend ourselves.
Go, unlike Chess, has deep mytho attached to it. Throughout the history of many Asian countries it's seen as the ultimate abstract strategy game that deeply relies on players' intuition, personality, worldview. The best players are not described as "smart", they are described as "wise". I think there is even an ancient story about an entire diplomatic exchange being brokered over a single Go game.
Throughout history, Go has become more than just a board game, it has become a medium where the sagacious ones use to reflect their world views, discuss their philosophy, and communicate their beliefs.
So instead of a logic game, it's almost seen and treated as an art form. And now an AI without emotion, philosophy or personality just comes in and brushes all of that aside and turns Go into a simple game of mathematics. It's a little hard to accept for some people.
Now imagine the winning author of the next Hugo Award turns out to be an AI, how unsettling would that be.
We still have chess tournaments, super-star grandmasters and circus freaks (people who can play blindfolded against multiple opponents). And, yes, computers can easily smoke all but elite players.
Why should Go be different?
And yes, chess has lost some reputation. And I guess it will be similar with go. I mean there is a University just for go. But learning something where you know you can become the best, is something different, than learning something knowing computers will be allways better than you ... so I guess they are having a hard time right now ..
FTFY
In chess the top engines are rated hundreds of ELO points above Magnus Carlsen (top human). No top ranked human vs computer match has been publicised in over 5 years because humans are thoroughly trounced. There are cyborg matches which are interesting. Human + Computer vs Human + Computer because gameplay techniques are considered different. Humans still depend more on higher level goal strategy and less on ruthless positional efficiency (which is probably why they get beat midgame).
What is mind boggling is that 6 months ago no go engine was scratching the surface of professional level go. It took the engines getting a 4-5 stone handicap to be competitive at the lowest level of professional levels.
It looks like this one algorithm has blown through the professional ranks in about 3 months. And a 5-0 victory here would be like 2006 vs 1996 (or even 1993) chess in 3 months.
As for Go, I guess I would never had made a long bet against computers. But as recently as just over a decade ago, computers lost to merely competent players and people working on Go programs were pretty much saying that they didn't even know what the path forward looked like. Things improved a lot with Monte Carlo but even that stalled out. Admittedly, I don't follow this area closely, but these wins pretty much came out of nowhere.
(Not saying this is a bad thing. Evolution in games is natural, and I think it's amazing how much innovation is going on right now (particularly enabled by Kickstarter) - you'd think that board game design would have been worked out decades or centuries ago, but in the same way that incandescent bulb development accelerated massively when competition arrived, it feels like game design has got so much better when forced to compete with computer games. If there are other activities that people find more fun than Chess, that's all to the good)
I also played chess as a kid and the allure of both local and national tournaments was that you could play with a multitude of different players, as opposed to the same 4 or 5 habitual chess players in your family/school/circle of friends.
But now, with the internet, at any second you can play with different people from all over the world, different strengths, styles and whatnot.
Hence now, instead of looking for the local chess club in the weekends we can play, any time of the day, any day of the week, anywhere.
Sites like the excellent lichess [1] are even free (in this case, free both as in beer but also as in speech) and, at any moment there there are 9 thousand, 10 thousand players enjoying this magnificent game.
Right?
The way it picks moves is very similar to how top professionals do.
Intuition is reduced to memories stored vaguely as neural connections.
They can see certain stats and high level overview.
But they have no idea what it's thinking.
They know the general pattern of the algorithm. They even explain it.
But the algorithm involves two deep neural networks, and they don't really know what's going on inside them.
One of the developers showed up during the commentary on the second game and talk about this stuff:
Well apparently my question leads to some deep mathematic theories about languages encoded by data. http://cstheory.stackexchange.com/questions/15039/why-can-ma...
So the most likely explanation is that policy/value nets in AlphaGo have learned to extract - with cold logic - the key factors that make up what humans believe to be "good" board positions.
It has little to do with voodoo about neural connections and magic emerging from the weights. AlphaGo has most likely managed to identify the important factors of good board positions (by seeing tons of examples of good and bad moves/positions). It only appears to be magical because these factors are most likely very complex and inter-dependent.
This is supported by the AlphaGo paper - they report that AlphaGo without tree search is about as good as the best tree search programs (amateur pro level). So AlphaGo has taken amateur-pro-level board analysis ability, and combined it with tree search, to achieve top-player performance..
I don't think that's an entirely accurate way of putting it, because a player who reaches that level is also doing a little tree searching. Maybe if you found some human who managed to reach "amateur pro" level by playing purely on snap, instinctual decisions without any logic or exploration of variants at all, yes, then you could say their ability to evaluate positions is as good as AlphaGo's.
But I would guess that you are right anyway that we can deduce that its ability to evaluate a board position really is below that of better professionals, and its huge strength is due to the tree search (which of course involves a second "policy" net to pick moves to explore).
Isn't that pretty much the definition of intuition? Combining a bunch of things in some unknown and nonlinear way to result in a 'feeling' about the situation?
Isn't most human experience an overly mystical interpretation of physics/chemistry?
The point is, the programmers don't quite understand what exactly are those features that the neural net is seeing.
"As a casual player of Go myself, some of the moves that AlphaGo made were crazy. Even one of the 9th Dan analysts said something along the lines of 'That move has never been made in the history of Go, and its brillant.' and 'Professional Go players will be learning and copying that move as a part of the Go canon now'."
http://gobase.org/replay/?gam=/games/inter/matches/AlphaGo/L...
One interesting thing that happened during the time for Sedol's next move was that the 9th dan commentator started referring to AlphaGo as "he".
Other than the kake(shoulder hit on the right) the game might have been a regular top-prop game.
It's an upright denial to the way of life they so chose and devoted.
IMHO Google should donate the prize towards Go education and Go organizations instead of some random charities.
I don't think people will pay to watch Go Bots square off, but I think this example of "obsolete education" is a great reminder that it's not just the assembly line jobs on the chopping block.
So they're doing what you want them to (I can't find a summary of how they're allocating the money across each category). Personally I think the work UNICEF is doing to help women in developing countries is more important than Go charities, but I guess their choices should satisfy everyone.
And the fact, that, prior to the deep learning revolution, Go is the only board game that human cannot be beaten, add even more myth and charm to the game and players alike.
Now, it comes to the time, that Go can be modeled by computers, and hundred years of human study is topped by computer in less than a year's time. All those myths around it will be gone. That is the biggest bummer I guess.
And who can make the most interesting new go-like game?
Perhaps this could be tested with chess or checkers, even.
If an AI won the next Hugo award, I would be rejoiced. It wouldn't mean the end of literature at all; it would mean that humans are ready to produce an even higher form of literature.
https://en.m.wikipedia.org/wiki/Emily_Howell
We have all kinds of visual art made by computers and AI - prom painting from photos to abstract art to 3D renders.
We have computers writing poems and haiku.
The only thing that's missing is the conceptual creation, which, let's be honest, most human artists struggle at as well. So writing and interesting story is not yet in the AI's domain.
https://en.wikipedia.org/wiki/David_Cope#Discography
And they were well accepted by the music community.
I'm sure soon (if not now) AI can easily create art that regurgitates popular trends in the past, and perhaps some artists may find a way to use AI / other algorithmic techniques in a way that complements their personal vision. But AI is a long way off from replicating the quirks of human nature, the unique personalities and personal visions of humans. Until that happens, I can't see terribly interesting art emerging from AI alone.
Literature is not a lower form of art that we must strive to automate so that we can dedicate ourselves to more "complex" forms.
You are confusing the unknown with art.
> If an AI won the next Hugo award, I would be rejoiced. It wouldn't mean the end of literature at all; it would mean that humans are ready to produce an even higher form of literature.
To me this seems to be claiming that what we have now is a form of "lower" literature, to be tackled by AI so that humans can produce "an even higher form of literature". But, of course, literature isn't graded in a scale of "low" to "high". (Well, there is lowbrow and highbrow, but that's something else).
The mention of medicine as "holy art turned into boring science" (already somewhat dubious) also seems to point to the idea that it is art that's being "solved". But I admit I might have misread it.
By the way, I don't rule out that art can be produced by an AI (whatever that means). I subscribe to the notion that art is in the eye of the beholder, so if humans can find meaning in something produced by a non-human, that's probably valid art!
Being "low" or "high" is all dynamic. We already have a good example: the advertisement industry. When a way of advertising your product first came out, it is fresh and captures people eyes. As more and more advertisers follow suit, it became bad ad, and advertisers are forced to find new ways to attract people. Basically the criteria for good ads changes all the time, but that doesn't kill the ads industry.
Now imagine if AIs can write sci-fis that are "good" according to today's criteria. That would mean there will be loads of "good" sci-fis in the market, and people soon get tired of it. Now sci-fi authors have to come up with more creative ways of writing good sci-fis.
So AIs being able to produce literature means more variations and faster iteration in literature style, much like the ads industry today. I don't know whether this is a good or bad thing, but it is certainly far away from the death of literature.
I'm specifically objecting to your notion of art.
The advertisement industry is not a good analogy. It can indeed be improved, possibly by automated means. In contrast, the progression from "good" to "better" art doesn't work like that -- if it even exists at all! What is your measure of quality, anyway? Complexity? But sometimes minimalism is preferred in art. Maybe how many people like it? It doesn't work either; a lot of people like stuff that is not enjoyed by the majority.
When is art "better"? How can it be "improved"?
PS: the Sci-Fi market is already flooded by below-average human writers, so we don't need an AI to picture this nightmare scenario of good SF writers struggling to sell their books :P
This can be claimed to be true when we understand how deep neural networks mathematically.
I think that would be pretty awesome and amazing, to be honest.
Imagine the best book you've ever read. Entrancing, enlightening, cathartic. You reach the end, and it's ... perfect. Oh hey, a sequel. Wow, the sequel is just as good as the first book. It expands upon it without diminishing the original -- you feel better, more complete for having read it. Wait, is that a third book in the series? Wow, it's even better than the first two! A fourth -- well, maybe you should go to work now, it's Monday, but the book is so good. Calling in sick once won't hurt anything.
Imagine a perfect series of books, published without end, each better than the last, a new one coming out weekly ... daily ... hourly ...
Scary :)
The practical problem with this is that, as I understand it, the deep learning system needs a pretty large data set to work with to infer rules from. You can do this with go because there is a constraint on legal moves and a deterministic win condition, but given how vast the number of potential novels is (If we count the space of all ten thousand word collections of grammatically acceptable sentences) the existing number of novels may no be enough to infer a pattern. (Though possibly you could split the problem up by separately doing the natural language processing and abstractin out the plot)
In the end all life is is one choice after another, and making good ones over bad mostly leads to a happier life.
Well, I am of the opinion that mathematics is the language that subsumes all other kinds of languages and line of thoughts. In the end we shall be able to describe every idea or thought in purely mathematical form.
That is a very ironically imprecise sentiment.
Just wait until machines start producing top notch research in experimental fields such as chemistry and Physics...
I've been thinking precisely about that. I think a book written by a machine will make the NYT bestseller list within our lifetimes (I would give it a 75% chance within 10 years, but that's just a gut feeling).
I also think my kids will live long enough to see an animated movie that is conceived of, written, scored, and animated by an AI.
But generating a coherent narrative, and good writing to "implement" that narrative. These are huge problems which - as far as I'm aware - would require major breakthroughs to achieve. Machine translation is still utter garbage, and that's fairly straightforward work. We're nowhere near an AI which actually understands language.
Sure. I think they will be solved in the next 100 years though.
i believe all these ancient arts are making a big comeback. the roots are still there, they just got concealed over the centures.
Because strategies like "ripping out someone's intestines" are illegal, boxing and wrestling have an unfair advantage because they don't have to worry about that stuff to begin with. For more realistic fighting situations, see: https://en.wikipedia.org/wiki/Lei_tai
An outsider, new to the game, had managed to pick up and challenge top players successfully in a venerated game.
The reactions of community to this is uncannily similar.
Considering some of the recent Hugo winners, it turns out the story doesn't actually have to be good, so yeah, a computer-written story winning is probably closer than we think.
While I'm not able to comment on the length/depth of history of chess vs. go, the above statement seems foolish. Chess also has a lot of mythology and mystique attached to it. Champion chess players (perhaps more so a decade or two or three ago) are also treated with a respect that is not casual.
Why do you think AI has no emotion, philosophy, or personality? We too are mere machines. Magnificent machines, no doubt, but machines nonetheless.
There's 10^50 atoms in the planet Earth. That's a lot.
Let's put a chess board in each of them. We'll count each possible permutation of each of the chess boards as a separate position. That's a lot, right? There's 10^50 atoms, and 10^40 positions in each chess board so that gives us 10^90 total positions.
That's a lot of positions, but we're not quite there yet.
What we do now is we shrink this planet Earth full of chess board atoms down to the size of an atom itself, and make a whole universe out of these atoms.
So each atom in the universe is a planet Earth, and each atom in this planet Earth is a separate chess board. There's 10^80 atoms in the universe, and 10^90 positions in each of these atoms.
That makes 10^170 positions in total, which is the same as a single Go board.
Chess positions: 10^40 (https://en.wikipedia.org/wiki/Shannon_number) Go positions: 10^170 (https://en.wikipedia.org/wiki/Go_and_mathematics) Atoms in the universe: 10^80 (https://en.wikipedia.org/wiki/Observable_universe#Matter_con...) Atoms in the world: 10^50 (http://education.jlab.org/qa/mathatom_05.html)
I wouldn't go with most (because I don't know about that), but many of these boards would also be either impossible to achieve (in a normal game) or illegal.
Evaluation function exists but it is not as simple as it can be for chess.
Are they not? MoGo beat pros of 9 Dan on 9x9 in 2011: https://www.lri.fr/~teytaud/mogo.html
Fuego beat a pro in 2008 using MCTS actually.
Not sure what you meant regarding MCTS, I never said anything about MCTS not being able to beat pros.
See, a chess program needs to find a lot of valid moves (see Deep Blue which won because it had stupid but extremely fast HW move generators), evaluate the moves and do a very deep search, up to 14, out of the very few alternatives. Russian chess programmers were better those times. They came up with AVL trees e.g. But hardware won.
In Go it's completely different. A move generator makes no sense at all, and a depth search of 14 neither. There are not a few alternatives, there are too many. What you need is a good overall pattern matching of areas of interest and an evaluation of those areas. And we saw that this feature outplayed Lee Sedol. Sedol couldn't quite follow in the recalculation of the areas.
Same as in chess AlphaGo learned the easy thing, that the center is more important than the corners, something Lee forgot during the game. But it's not a deep search, it's a very broad search, and very complicated evaluation function. A neural net is perfect for this function.
> whereas in Go no such function seems to exist.
It does exist. It's the neural net. It's a simple pattern recognizer, which learns over time more and more.
I should also say that it's somewhat clear that Sedol made one suboptimal move, and AlphaGo capitalized on it. Interestingly, the English commentator made the same mistake as he was predicting lines of play. This involved play in the center of the board, in a very complicated position. Prior to this set of moves, the game was almost a tie. Afterwards, it was very heavily in AlphaGo's favor.
Does this mean AlphaGo was playing at a higher level or is it just a coincident?
Myungwan Kim 9-dan professional (comments on match #1): "She knows everything [...] Unthinkable move for human [...] AlphaGo plays like the god of go"
I want to see AlphaGo vs AlphaGo :-)
One benefit of programs like AlphaGo eventually becoming available to individuals is that we'll see a rise in some incredibly strong young players that leave their ancestors in the dust, as I think has happened with Chess via Magnus. For the amount of training that can only be done by playing more and more games, being able to do that against a computer will be a lot more efficient.
With chess, there were two breakthroughs. First, there was Deep Blue, which threw massive hardware resources at the problem and achieved world champion level play.
That was interesting, of course, but didn't really do anything for human chess, because most humans did not have access to the necessary hardware.
The second breakthrough was when the developers of chess programs that ran on commodity desktop computers improved their algorithms to the point that they could play at (and far beyond) Deep Blue's level even though they were only able to search about 1/100th as many positions per second.
That was when humans started being able to really use computers to help the humans understand chess.
The breakthroughs in chess algorithms on commodity computers had little, if anything, to do with the Deep Blue breakthrough. The two are just too different.
Can AlphaGo be made available on hardware that top human Go players have access to, or is AlphaGo to Go as Deep Blue was to chess?
Whether those insights will come soon or not is the big question.
I don't think that's actually true. The hardware that AlphaGo is on is probably a lot more powerful than the one that is available in a single human brain, the big difference is in the software.
See the difference between the very best chess programs of a decade ago versus the ones now.
Its somewhat interesting to think about the differences in marketing between IBM and Google; IBM was marketing hardware and HPC with deep blue, but Google is marketing AI when so much of their advances in AlphaGo are enabled by distributed systems and HPC running billions of games training deep neural networks. It feels a little smoke and mirrors which is probably why they won't release much until after they get enough marketing value from this tournament :)
Can you please indicate this time in the video of the game? Thanks a lot.
It was fairly early in the game and about 4-5 lines down from the top and towards the center, center-left. Apparently, Lee Sedol played a little conservatively and "didn't take a ko" (?) when he could have.
Hope that helps you figure out what they're referring to, or maybe someone else can chime in, but that's what I remember.
You try and share this story with a non-technical person and they will likely say "Well, duh..it's a computer".
Is it that the deep net works mainly as an evaluation function of the current position? I guess it does more than that right?
Chess has a fairly straight forward ranking of pieces, fairly well established ranking of piece power. A knight is more or less always a knight.
Go, you kinda make up your "pieces" from scratch and there isn't any single common ranking. This makes the problem of even evaluating the board difficult (value), which is a sub step of making your higher plans (policy).
The output of the two dNN help cut off, or prevent the sampling of the bad game states and encourage looking into the fruitful areas.
Here is a (edit: not old) new paper, https://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf
I mean, deep down, Go is a perfect-information-zero-sum-discrete-2-player game.
I think a more successful pitch is showing that the strategies and skills coming from Go (human or robot) can be useful outside the game, as I argue here: http://rare-technologies.com/go_games_life/
As Radim says, of course, the intuition is only relevant when logic fails. However, no computer, including AlphaGo, has sufficient processing power to take the logic-only approach to the, uh, logical extreme (more board states than atoms in the universe, etc etc).
So both humans and computers play by a combination of logic and intuition. Surely Lee Sedol must be incredibly good at both. Perhaps AlphaGo is better than Lee Sedol at the logic part (or perhaps not, but just suppose for the sake of argument).
In this light, what is interesting is that AlphaGo is sufficiently good at intuition, a domain we might have considered uniquely human[0], to complement its ferocious logical power.
[0] I find it foolish to consider anything uniquely human, but a humanist essentialist might make such a claim.
I'm still holding out for the discovery of a proof of winning strategy like other combinatorial games. That to me is the logical extreme, not magic evaluation of all states. There are two interesting books on analyzing Go with combinatorial game theory, they found a neat system of scoring in the endgame and can create example boards with one best move that can be found by mechanically applying their system but stumps 9p players (who use whatever systems).
If software started writing bestselling novels, it would soon become a "duh that's just what computers do" matter of fact.
Here is what humans can do: When presented with pretty much any task, specified however poorly, they can design hardware and algorithms that can beat themselve at that task.
That's a decent measure of intelligence. A decent measure of creativity is coming up with tasks that make the intelligence part as interesting as possible.
But on one hand, in order to have a shot at passing the Turing Test, a computer has to first be able to understand and speak human language. And that's pretty damn hard actually.
Yes we can deploy statistical methods, deep learning or what have you for classification, feature extraction and so on, but natural language is context rich and ambiguous. NLP isn't a conscious process for us either, but our brains are capable to disambiguate effortlessly. And a big part of that is also the rich (human) experience we gain while growing up. Being able to speak natural language is the one big trait that separates us from animals. It's why we can cheat death, because our children can learn from the acquired knowledge of all of their ancestors.
This is different from playing Chess at least. With chess you can deploy smart algorithms, but in the end it's still a raw search for good moves in the space of all possible moves, while giving up on branches with a bad score. Raw search is what computers have always been good at. That's not really possible with NLP, because at least with chess your vocabulary is very limited and for calculating good moves you don't need extensive knowledge about the world we live in.
Computers will surely be able to do that in the future and that will be a major milestone towards "true AI", but for now computers cannot do it.
On the Turing Test, some features considered as being unintelligent, like the tendency for lying, or being sensitive to insults are actually evolutionary traits, that have arguably helped humans to survive. So while emulating some human traits will be counter productive, a "true AI" will be concerned with survival and as a consequence will end up doing whatever it takes.
So while passing the Turing Test may not be enough, not passing the Turing Test is a sign that the computer is unintelligent.
"im so bored this test sux", "i dunno wat are you asking me", "what if ur a bot lol"
AND THEN I WOULDN'T ACTUALLY KNOW because some people ARE this stupid
I know the kind of people you're talking about. I have a family member like that. She's not that smart, she failed her baccalaureate, she's semi-illiterate and she probably suffers from ADHD, though a lot has to do with her upbringing. But getting answers like that from such a person means that you're not asking the right questions or the incentive to answer is not there. Give a cash reward to a cash-strapped person and you will never get an answer like "im so bored this test sux".
Check out http://www.jabberwacky.com/
it jokes with you, it gives vulgar answers sometimes, etc.
it's NOT good enough to fool someone, but what if you just made it sound like it's a person with a disability? Is that still beating the Turing test? Is talking to a "five year old" still beating the Turing test?
Or does it have to be a 100 IQ adult person fluent in English? In which case that's just improving the bot a bit. To make it a funny and charming bot it would require even more effort. But it's just slowly improving the state of the art with some kind of techniques like reinforcement learning or neural networks or whatnot.
When a bot actually beats the Turing test it wouldn't be big news because it would just be a slightly better jabberwacky.
> So while passing the Turing Test may not be enough, not passing
> the Turing Test is a sign that the computer is unintelligent.
Similarly you can conceive of an AI that's smart, self-aware and intelligent just hasn't been developed to talk to humans.
The Turing test is a fine test to figure out if your AI is conversational with humans, but the OP I was replying to was suggesting it as a general AI intelligence test, it's not meant for that, and will give you both false positives & negatives.
The average person isn't impressed with computers winning at Go because they vastly underestimate the complexity and open-endedness of Go and wonder why it's really all that much more complex than chasing Pac Man through a maze like their computers were doing quite happily, and even with apparent personality, in the 1980s.
The computer was able to overcome the computational difficulty humans compensated with abstractions and strategic concepts. We still use those to understand whats going on in the game, but the computer is oblivious to them and only uses a tactical view of it.
It looks impressive, but I think any strong amateur can do that.
IIRC that's basic expectation of any Go player at a non-trivial level, starting from the mid-high amateur ranks.
It's completely expected that both players and observers can record the game on a kifu or replay and discuss it immediately after they finish.
Chess grandmasters can often play simultaneously and blindfolded against 20 different opponents (and win all but one or two games). This is nowhere near the extent of Redmond's memory strength. :)
Especially for Lee, the whole world is looking at him. An "ordinary" human like me won't be able to make the right decisions under this pressure.
A great respect to Lee and the Developers of AlphaGo. Good Game!
There's just no useful way to "pool" human thinking power and redistribute it to where it's needed the most – all the players will simultaneously consider the same branches. At best you reduce the risk of making a silly mistake, but sitting pros are already pretty good at that.
A computer-assisted pool of humans might work. Feed each human a board state advanced by 1/2/3/N moves down the decision tree in some direction, and have them evaluate that particular sub-tree. It's a map-reduce problem!
Are Bridgewater Capital's employees Ray Dalio's focused?
https://en.wikipedia.org/wiki/Quarantine_%28Greg_Egan_novel%...
Is there a meaningful response that each one could give that would be optimal? I figure this would only stand a chance of working if the pool of humans were cloned from a 9 dan :-)
Of course, Go has way too many eligible possible moves for random people to vote on, but a large enough group of top pros might be able to do well just by voting.
Four to five expert chess players suggested moves for the world team. I feel that for any reasonable non-expert it comes down to "choose between these suggested moves" rather than "pick a move". It's also not really random human beings, since the selection of participants is self-selected and therefore much more likely to contain very good chess players.
This is mostly semantics, but anyone who doesn't read your link might get the wrong idea, so I felt like clarifying it a bit.
The situation with Go is different. (I wrote the Go program Honninbo Warrior in the 1970s, so I am a Go player and used to be a Go programmer.) Still, I bet the AlphaGo, and future versions, will strongly impact human play.
Maybe it was my imagination, but it sometimes looked like Lee Sedol looked happy + interested even late in the two games when he knew he was losing.
EDIT: clarified to what I originally meant: "end of midgame"
I'm reminded of the time when Deep Blue made a buggy move that totally threw Kasparov off because it was so counterintuitive. But it turned out to be a real mistake. If Kasparov had kept his composure, who knows what could have happened?
http://www.wired.com/2012/09/deep-blue-computer-bug/
Of course, as been stated here, AlphaGo doesn't have the issue of having to keep composure, while Lee does have that issue and pressure. But who's to say that AlphaGo won't make a similar significant mistake, and that Lee won't be able to capitalize on it to make a comeback in that particular game?
edit: admittedly, as pshc as noted, the search space probably grows smaller as the game continues.
edit 2: suddenly, his comment disappeared for some reason?
I'm not saying you're wrong, I'm just saying I don't agree with your motivation.
In any case, I personally find it more interesting to see what we can learn from AlphaGo about the opening and the midgame.
Where I wrote that a pro will blunder "now and then" I should have written "rarely" -- I don't have data to back this up but I'd guess once in every n games for some n > 30.
In particular, I'm wondering if a computer scientist with access to the alphaGo source code and all the weights of the network could trick alphaGo in order to win games automatically (cf. the papers that show a neural net can be tricked to classify a plane as any other class).
If a human with the knowledge of the source code and the weights can do this, it is scary. Imagine a similar algorithm runs your car. An attacker that knows the source code and the weights may trick the algorithm to send your car in a wall!
Tricking the NN works in a no noise situation, also an image has many more parameters than a Go board
I'm kind of hoping for an unconventional opening in game #3. Come on tengen or 5,5... Additionally I think that might be one way of weakening the bot in that it will find its net for suggesting candidate moves less useful and lean more on unguided MCTS, but that's just a wild guess.
Though I wonder if something emerges. For example, black always wins, or, black always wins from a certain opening position.
Great game btw, a pleasure to watch.
In a sense this is unfair: alphaGo was trained with a lot of human data, but AFAIK Lee Sedol is playing AI for the first time.
I would assume the development will continue.
However, they won't make it such a big PR focus. (And even if they stop, other people will implement the same ideas.)
http://www.shanghaidaily.com/national/AlphaGo-cant-beat-me-s...
Ke Jie believes he is slightly better than AlphaGo given the play he has seen, although he thinks AlphaGo will outpower him in a few months.
The belief that every good player that has so far played against AlphaGo would win highlights a common misunderstanding of the nature of the human mind. There seems to be in fact a very small difference between the aptitudes of the average players and those of the top players, such that it is very easy for a machine to go from average to superhuman. Getting to average is by far the harder part.
This is an interesting observation across many domains that involve "intelligence". It's actually pretty easy to see why, if you reconsider the scales for "good" vs "bad": The "bad" end of the scale is not at "the village idiot". The "bad" end of the scale is at "a rock". So, yes, once you get from "rock" to "village idiot", "superhuman" is a comparatively tiny step away.
Closing in on nearly 600 career championship wins.
https://en.wikipedia.org/wiki/Michael_Redmond_%28Go_player%2...
Chris Garlock, on the other hand, doesn't add that much value to the broadcast. Maybe somebody will start a "Left Commentator" meme, just like Left Shark.
When searching the game tree, at each ply the most promising N moves are examined (as determined by the policy network) and leaves of the game tree are scored by the value network.
Nonetheless, AlphaGo takes a minute and a half to play its next move. Can anyone explain what on earth is going on during those 90 seconds?
A person might memorize standard openings and play them by rote. The computer _could_ do that, but it would just be a computational short cut. In a sense, the machine is re-discovering the classic opening.
You have to realise that it's not using an opening book, it's computing the best response from scratch.
You can hold out for a few thousand years, but eventually the uncontrollable and amoral technological imperative will catch on and crush you.
It's kind of poetic and sad. It feels like technology will render everything un-sacred eventually.
good to know they'll play all 5 games no matter what the result is though
People seem to think Lee knew he lost and was just playing to learn more. Hope he learned enough to take the overlord down in the next three games
I wonder if Lee Sedol will have an interest in studying deep learning after this =)
Both players get the same time controls, seems fair to me
But if you're saying humans might fare better against computers in a game with a longer time control, I suspect that's true
It seems that additional time might actually work _against_ humans and for AlphaGo.
The human's strength is intuition and insight. They can look at a move and have a good understanding of strengths and weaknesses of positions by "feel" developed by long practice of the game. More time doesn't really help this much.
Another part of the game is "reading" -- playing out scenarios of response and counter-response to evaluate how strong a move is. The computer excels at this, because it can play out as many moves as its computation time allows and remember all the results with complete accuracy. Humans are slower and prone to mistakes when they do lots of reading.
So adding clock time lets the computer increase its advantage over the human in reading depth, but doesn't so much increase the human's advantage of intuition and insight.
And you can always give the computer less time than the human, but this just shows that it's stronger than you and you need to handicap it to have a chance.
Increasing the human time is not an option, since no one wants to watch 8 hour games, and fatigue could also come into play, so the only real option for a winning chance is reducing computer time/power
Correspondence go, similarly, would see fewer errors. Holding a world championship would take quite a bit longer than in chess, though (rough guess: 300-ish half-moves per game versus 100-ish half moves, the latter, I guess, with a bit more variation). That could be problematic, as a world championship in chess already takes years (curiously, some championships finished before the one started a year earlier did)
FWIW regular tournament Go games go up to 6 hours, and title games can take more than 16h and span over two days.
Unless you're not counting support systems, in case it becomes very complicated to calculate and decide exactly which energy expenses are for support systems and which are directly integral for function.
Lee Sedol is vastly more efficient than the entire AlphaGo cluster. However, while AlphaGo gains a predictable amount of power as its computing power is increased, it's not clear that one could do the same with humans. Our Go players optimize individual play, not multi-brain distributed play. What would the match look like if we trained up a bunch of humans to play Go as a team, and pitted AlphaGo against a team of humans that consume the same number of joules over the course of the match as it does?
I admit my first thought on fairness did go in the same direction of limiting energy budgets. But after reflecting on it just long enough to realise the above, I am finding myself surprisingly uninterested. It now seems to me that nothing particularly insightful would be revealed: limiting energy budget is no less arbitrary than limiting time unless the artificial opponent is expected to be capable of a range of things comparable to that expectable of an average human. Or if expectations are much lower, the artificial opponent would need to contend with drastically tighter limits to approach “fairness” – though at this time it would be guesswork how much tighter they ought to be. Either way, it is glaringly obvious that no computer would come within miles of competing.
So ultimately the fact that Go has been “broken” (in a particular sense) at all is far more interesting to me than whether the machine is competitive with the human in any more general sense. “It’s not” as the universal answer is boring.
And to digress a bit from there: From that perspective, this was ultimately a very human achievement. It was humans who chose Go as a problem to attack and it was them who picked MCTS and deep learning as the way to go. (Uh, no pun intended.) That’s not just reassuring. It’s also a framing we should keep in mind as computers become more entangled with the physical world and more autonomous.
Tonight's game was beautiful. Last night's was a fighting game way too high level for me to really grasp (I have no idea how to play like that, all those straight and thin groups would make me nervous). I'm expecting Sedol to win Friday since I imagine he's going to have a great study session today, but I'm no longer confident he'll win the last two.. Still rooting for him though. :) (I also want to see AlphaGo play Ke Jie (ed: sounds like from the other submission on Ke's thoughts that may happen if Sedol is soundly defeated), and for kicks play Fan Hui again and see whether it now crushes weaker pros or is strangely biased to adopt a style just slightly stronger than who it's facing.)
Not sure if any of the games he played are available, though.
1. As the game of go progresses, the number of reasonable moves decreases, so that as the game progresses, players on average play closer and closer to optimally. By the end of the game, even weak amateurs can calculate the optimal move. Logically, I would guess that stronger players are able to play optimally earlier than weak ones. Lee Sedol is known for his strong middle and endgame, often falling behind early on and making it up late in the game. He is so strong at this that he has driven an entire generation of go players to developing very strong endgame. But AlphaGo, running Monte Carlo simulations, almost certainly can brute force the game earlier than Lee Sedol can. Lee Sedol is playing AlphaGo on its own turf. A player known for their opening prowess, such as Kobayashi Koichi in his heyday, might have had an advantage that Lee Sedol doesn't. (Note: I'm not strong enough to analyze Lee Sedol or Kobayashi Koichi's play styles; I'm repeating what I've heard from professionals.)
2. I hoped that when an AI beat a pro at go, it would be with a more adaptive algorithm, one not specifically designed to play go. If my understanding of AlphaGo is correct, it's basically just Monte Carlo: the advances made were primarily in improving the scoring function to be more accurate earlier, and the tree pruning function, both of which are go-specific. It's not really a new way of thinking about go (at least, since Monte Carlo was first applied to go). It's just an old way optimized. The AI can't, for example, explain its moves, or apply what it learned from learning go to another game. It's certainly a milestone in Go AI, and I don't want to downplay what an achievement this is for the AlphaGo developers, but I also don't think this is the progress toward a more generalized AI that I hoped would be the first to beat a professional.
The particular algorithm used by AlphaGo is of course specific to Go (the neural network inputs have a number of hand-crafted features), but the overall structure of the algorithm - MCTS, deep neural nets, reinforcement learning - is very general. So there's two ways to look at it. One is that what you wanted has actually transpired.
The other is that what you asked for is completely unreasonable. I think it highly unlikely that an algorithm not specialised to Go will ever be able to beat all specialist Go playing programs.
AlphaGo can't explain the outputs of its two NNs, but it can still explain its moves by showing which variations it thinks are likely.
It is general in the sense that humans can apply those algorithms to different problems (and have been doing so for decades). It isn't general in the sense that we can't apply AlphaGo to other problems unmodified. AlphaGo can't even play chess badly. It is not really even a step toward strong AI. (Note that "strong AI" is a term with a specific meaning. [1])
> The other is that what you asked for is completely unreasonable.
That's tantamount to saying strong AI is unreasonable.
No, what I said is that it's unreasonable to expect that strong AI would play better Go than whatever the contemporary state-of-the-art Go AI is. But stated like that, I'm not sure I can agree with my statement. Strong AI could design and implement its own specialised Go AI. How would you count that?!
I think we might not be able to answer these questions until a strong AI emerges.
Personally I think the approach of combining deep learning with MCTS to beat Go was obvious to anyone sort of familiar with each thing, and with a good funded team could be done in a year or less, but a lot of 'experts' were ignorant of one or more of the areas. The uncertainty should have revolved around when some group would get around to writing the software and scaling up with powerful hardware. Implementation details, the theory work was already known.
My own lesson from this is that if all that is needed is tough engineering work (but not really new theory) that work could literally arrive in a week instead of the x+delay time it might take from scratch, because some group could already have been in the process for about x time that isn't public knowledge. AlphaGo kind of came out of nowhere; there were early signs with papers on using deep learning techniques, but I don't remember any public commitments to much. That just indicates companies are still quite capable of doing secret projects. If they had a secret new theory, too, it could be even more amazing. I know of one startup in particular that's been securely at the top of its niche because internally they have secret CS research unknown to academia.
The OpenAI initiative may be useful from the perspective that if they've shared what probably shouldn't be shared, at least we can suspect something is imminent and try to plan a last minute stand of getting it right first, vs someone like Google doing all the research in secret and then bam, unleashing UFAI.
Especially when AlphaGo capitalized on just one suboptimal move of Lee Sedol.
This is a milestone in modern informatics.
As of 2015, the best player in the world is a computer.
There is something unnerving about a computer that can answer in 0.01 seconds and still have the move be better than any human would come up with in an hour. At that point a robot playing simulatenous bullet chess would wipe the floor with a row of grandmasters, beating them all without exception.
https://web.archive.org/web/20160128151110/https://storage.g...
I'd be interested in how strong it would be if given the same constraints as human learning (playing thousands of games, rather than millions).
Then suddenly a computer comes along and takes that title from you. But it takes it in such a way that you are never in your life able to re-take it because of how the AI works.
A game will likely just be the first field. My girlfriend is working in translation and interpretation which is another area already in the crosshair of neural networks. AIs will step by step become more efficient than people and that is terrifying.
Now, I'm not good enough to evaluate if this seems likely, but I could understand that AlphaGo (especially if it was winning) avoided Kos, since they are high variance situations.
I am very very very far from a Go expert, but Ko did always seem to be "first person to make a mistake looses". If that's true, I'm sure Sedol deliberately stayed away from it as an area where the computer is likely to be particularly strong.
So if the first comment in this thread (about how it's a completely non-human approach) is true, it's really interesting that humans can enable computers to come up with non-human ways of solving complex problems.
Seems like a big part of this story, if I'm not being completely dumb here.
Slack channel for discussion if anyone's interested. We're using it for commentary while the games go on. Was created by AGA people.
But AlphaGo showed us what AI is really capable of doing in an eerie sort of way and I think interest in AI will soon become mainstream which is a good thing for the development of AI.
Now it's at least easier to comprehend the context of all those doomsday warnings about AI destroying humanity which I never took seriously.
I'm looking at the DeepMind channel on Youtube: https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A
Last night I found that on the DeepMind channel they were explaining the game a lot for beginners. This was a bit tedious for me but then this is a historic event so there'll be noobs tuning in so I can understand that, and I appreciate that the TV people are thinking of them.
Also, the AGA channel you just seem to have a split screen showing two people with headphones on talking over Skype? (Granted, one of them is Myungwan Kim, 9p, a super likeable guy - but Michael Redmond, 9p, is also super likeable.) The DeepMind channel switches between the game board and the large commentary board, with occasional shots of Lee Sedol and occasional shots of the DeepMind terminal.
Michael Redmond was reading out lots of variations and repeatedly trying to count the board and evaluate the position. I think he was trying to be as informative as he could. Occasionally he would get lost in thought as he calculated owing to the complexity. You could call it many things but uninformative is not one of them.
edit: added pro commentators names...
Yeah, I think I caught some 15 minutes of this and stopped watching, maybe it got better later? I would have expected them to the introduction for beginners before the start of the real game, and they ended up skipping commentary on what seemed like pretty key choice in the early game to do so.
> Also, the AGA channel you just seem to have a split screen showing two people with headphones on talking over Skype? (Granted, one of them is Myungwan Kim, 9p, a super likeable guy - but Michael Redmond, 9p, is also super likeable.)
There were even some technical difficulties with the skype feed on a few occasions and I wouldn't call it great quality, but on the other hand I rather prefer the KGS virtual board - on which it's quicker to show the possible alternatives and easier to point out the key stones.
I switched to KGS at intervals to see what people were saying on the mirrored board kibitz. And the DeepMind commentary did begin to irritate at times but I feel they had to cater for noobs. Once the middle game got going it was fine though. Agreed on the KGS virtual board.
Until you are close to late-game the scoring of many of the positions is quite flexible (depending on sente, and how certain local situations are resolved) so although you would may have a semi-reasonable approximations for total points, the games have been close enough that a +/- 5 point error in estimating will make it impossible to tell who is actually ahead.
That said, it would be very curious if we could get a virtual-board with AlphaGo's evaluation function used to score each position.
I wonder what would be interesting games (intellectual sports) where computers have yet to defeat humans that you would probably be interested in learning?
What happens if the Go master tries to deceive the oponent? As in purposefully play a counter-intuitive position, or even "try to lose"? Will the AI's response be confused as it is expecting rational moves from its oponent?
"It's just like parents bathing in the success of their offspring."
There is no offspring here. There is a set of calculations, written by some people.
AlphaGo is not a child. AlphaGo is a set of deterministic calculations, defined and created by some people. Your question about humans and their children is irrelevant. There is no child here.
In this flawed analogy, the "parents" are the developers, and the "child" is a set of calculations. If someone wrote down a set of calculations, and then executed those calculations and won, is it right to say that the person who wrote down those calculations won? I suggest that it is.
I'm pretty sure that if I wrote a child from scratch (AGCTTAACGGUAA ... etc), understood the underlying mechanisms connecting proteins to wining a baseball tournament ... I should get some credit for the win :)
There is no insight in making a child (can be done totally drunk and half passed out), although there is some in education.
This has a powerful consequence: we have not seen AlphaGo pushed to the limit, he is lowering the distances as if it were playing a teaching game.
Lee Sedol I think came to this conclusion, and the only human strategy left is to take a lead big enough to maintain the rest of the game. And that might be the last strategy to play to show the computer is already unbeatable, because it will be pushed to its limits to win a game and it might overcome humans.
Some of the moves AlphaGo played both in this game and the previous one are very definite mistakes, but they were irrelevant to the difference in the game.
It is only a mistake to your human biases. AlphaGo literally has no conception of the margin of a win. It doesn't care either way how many points it wins by, as long as the win percentage is maximized.
When they interviewed the devs briefly, they said this is because AlphaGo doesn't really consider the score other whan winning, so it will pick a move it think is an 81% chance to win by one stone over an 80% chance to win by 10 stones. When it's ahead these moves can look like mistakes but a better way of describing them would be hedging.
"The idea of using ConvNet for Go playing goes back a long time. Back in 1994, Nicol Schraudolph and his collaborators published a paper at NIPS that combined ConvNets with reinforcement learning to play Go. But the techniques weren't as well understood as they are now, and the computers of the time limited the size and complexity of the ConvNet that could be trained. More recently Chris Maddison, a PhD student at the University of Toronto, published a paper with researchers at Google and DeepMind at ICLR 2015 showing that a large ConvNet trained with a database of recorded games could do a pretty good job at predicting moves. The work published at ICML from Amos Storkey's group at University of Edinburgh also shows similar results. Many researchers started to believe that perhaps deep learning and ConvNets could really make an impact on computer Go.
...
Clearly, the quality of the tactics could be improved by combining a ConvNet with the kind of tree search methods that had made the success of the best current Go bots. Over the last 5 years, computer Go made a lot of progress through Monte Carlo Tree Search. MCTS is a kind of “randomized” version of the tree search methods that are used in computer chess programs. MCTS was first proposed by a team of French researchers from INRIA. It was soon picked up by many of the best computer Go teams and quickly became the standard method around which the top Go bots were built. But building an MCTS-based Go bots requires quite a bit of input from expert Go players. That's where deep learning comes in.
...
A good next step is to combine ConvNets and MCTS with reinforcement learning, as pioneered by Nicol Schraudolph's work. The advantage of using reinforcement learning is that the machine can train itself by playing many games against copies of itself. This idea goes back to Gerry Tesauro's “NeuroGammon,” a computer backgammon player that combined neural nets and reinforcement learning that beat the backgammon world champion in the early 1990s. We know that several teams across the world are actively working on such systems. Ours is still in development.
...
This is an exciting time to be working on AI."
In future, it will be interesting to see AlphaGo playing against itself!