Slashdot Mirror
Google's AlphaGo AI Beats Lee Se-dol Again, Wins Go Series 4-1 (theverge.com)
An anonymous reader quotes an article at The Verge about Korean grandmaster's fifth and final game with Google's AlphaGo AI: After suffering its first defeat in the Google DeepMind Challenge Match on Sunday, the Go-playing AI AlphaGo has beaten world-class player Lee Se-dol for a fourth time to win the five-game series 4-1 overall. The final game proved to be a close one, with both sides fighting hard and going deep into overtime. The win came after a "bad mistake" made early in the game, according to DeepMind founder Demis Hassabis, leaving AlphaGo "trying hard to claw it back."
Nobody is trying to make a general intelligence because nobody wants it. What is wanted is domain specific algorithms that are very good at what they do.
Although, it seems that the tech is quite general and learned to play multiple Atari games without having to be tuned for each.
It's a huge surprise.
Your comment is a perfect example of the AI Effect:
https://en.wikipedia.org/wiki/AI_effect
Most likely. But that's the case with just about any game.
Can you really guarantee to find fair games of anything from tic-tac-toe, to chess, to draughts, to reversi, to Risk, to poker, to anything at all online? Go was pretty much out-there on its on in this regard but ordinary PC Go software has been able to beat amateurs for a few years now. This is orders-of-magnitude in terms of a leap in capability but nothing that would change the situation for the majority of people playing it.
Pretty much the only "games" that can be fair are if you can guarantee it's against a human without any kind of possibility they could be plugging moves into a computer at any point. That's a vanishingly small amount of plays, pretty much limited to strict competitions (and even professional chess competitions have seen people use toilet breaks to illicitly get computer analysis of the board state on their phones!).
What I want is not a computer player that never wins, nor one that wins all the times. Those are EASY to program in comparison to one that CONVINCINGLY challenges you enough that you have to play slightly better each time in order to win, without trouncing you or letting you walk all over it.
That's the REAL hard problem in any kind of game "AI" - the "gamer's Turing Test" - how to lose/win convincingly without people knowing you're a bot.
Try playing a pool game on a computer, for example. They usually go from "whoops, missed a blindingly obvious easy shot" to "four-cushion bounces, jump the ball, curve into the other balls coming back from the cushions, and tap one into the nominated pocket" without anything convincing in between.
Like Left4Dead's "director" - we need to adjust to the player just enough to make it fun but that if they're obviously letting their guard down, we take advantage. Unlike the 80's arcade games that had to be punishingly hard but just easy enough at first for you to want to put money in but not to waste too much time in front of, modern video games need to be easy enough to pick up and get you into and keep you coming back for more (and spending some DLC) without feeling like you're playing a script, trouncing everything, or need to spend a fortune just to stay competitive.
Actually, even this version will be better at the next game, as it is a self learning system. It became that good at Go by playing millions of games against diverse opponents.
Sorry, but I'm a mathematician. Check my comment history, I'm the first to disparage any kind of "AI" (which just means human-programmed heuristic most of the time), especially that which just does brute-force search of possibilities. That's NOT AI. Almost every "game AI" isn't AI. Not even close.
However, in uni, one of my lecturers was studying Go as one of his prime areas of research, and I've seen - and checked - some of the numbers here.
You have no idea what this machine has just done. It's leapt forward some 10-20 years in terms of computer Go-playing capability in one fell swoop. The numbers involved in Go are so huge that brute-force search, even for a limited number of moves, is absolutely impossible in the times given.
And it isn't being given programmed hints, because Go is just too complex a game for that beyond amateur play. There's a handful of hard-and-fast rules of what's a stupid move and what's not and everything else interacts SO MUCH with the rest of the board and future plays that it's almost impossible to even tell who's winning most of the time!
As such, this system, no matter the power behind it, is doing something that dumb, brute-force, play-the-game AI written by world-experts in Go, AI, and game theory wasn't expected to be able to achieve within the next decade. And it primarily gets there because it learns from information fed to it.
At that point, although it's only limited to Go, the engine is proving itself capable of - almost - a kind of intuition, insight and "feel" for the positions rather than anything to do with numbers and scoring and weighting and pre-written rules. Now, that's a vastly overblown explanation, still. The computer isn't "feeling" anything. But whatever it's emulation and use of such, it's leaps-and-bounds ahead of its competitors.
This is why it makes BBC News, Slashdot and every other media outlet. It's not just winning by brute force. It's doing something else. It's spotting patterns in data it's never been exposed to before. It's able to hypothesise and learn from mistakes on board layouts that maybe NO HUMAN HAS EVER SEEN BEFORE OR WILL AGAIN (that's how large some of the numbers of possibilities get!).
Even a pack of cards, with 52! = 8x10^67 potential arrangements of a shuffled deck:
http://www.murderousmaths.co.u...
Pales in comparison to the number of possible Go positions (2x10^170) and the ways that you can move from one to another (~ 1 x 10^768). And that's just on a standard 19x19 board (something almost unplayable for a computer just a decade again).
This thing isn't calculating. It's gaining insight from historical observation and applying that to self-similar situations that nobody has ever been able to analyse, nor which it could ever analyse fully in the time given. That's the start of "true" AI. It's only a start, but it's quite seriously ground-breaking in that ability.
And once you start down that route, there's nothing stopping AlphaGo quickly learning every similar game, then dis-similar games, then other games, then other things entirely, using the same kinds of system underneath.
Honestly, there's a reason that game theorists and AI-experts are making a fuss about this.
I've been following the matches with the same expectation and anger I felt in 1997 during the Kasparov & Deep Blue rematch. The final result has been similar, and although it has been well reasoned that chess and go are pretty different games and Deep Blue and AlphaGo are pretty different machines, the bittersweet sensation is identical. I had a naive hope in the human superiority just for a little more time. I was pretty sad after the final game: Lee Sedol seemed really disappointed and sad himself. I can't imagine the pressure he's felt throughout the event, and his face -that's my impression- seemed to tell us "I've failed you all". He later told in the press conference that he felt he could have done more in the games -I'm sure he'd like to play more games to test himself again- and I wonder what could have happened if the matches would have been played without general knowledge. Feeling that kind of coverage must have been really stressful. If you ever read this, Mr. Sedol, thank you. And please, don't ever feel disappointed, you've done a fantastic job.
Probably, but the fourth game was rather incredible and unlikely to be repeated again, even without changing alpha go. Lee took the corners and forced go into the much larger middle, Opposite of how he played game 2. This probably was the perfect setup to best alpha go, because it was too big an area to go deep into all possibilities of moves, yet to complex an area for the general strength of positions to be understood. They way the centre unfolded ended up a dream knife for Lee, Lee built to it when he saw it, but it emerged unexpectedly from either player and seems to be an incredibly rare number of setups where it would work out (wet know this because some lees moves before 70 could have been even better to take the middle). Lee was amazing to see it and take advantage of it, but you could probably attempt the same strategy against Alfa go a million times and taking the centre like that would never work.
No no, you're thinking is all wrong. In 10 years, we'll pitting our phones together on the same table and have them play it out while placing winning bets . It's sorta like putting two Furbies in front of each other; useless, but endless fun :)
Life is not for the lazy.
You have no idea what this machine has just done. It's leapt forward some 10-20 years in terms of computer Go-playing capability in one fell swoop. The numbers involved in Go are so huge that brute-force search, even for a limited number of moves, is absolutely impossible in the times given.
And it isn't being given programmed hints, because Go is just too complex a game for that beyond amateur play. There's a handful of hard-and-fast rules of what's a stupid move and what's not and everything else interacts SO MUCH with the rest of the board and future plays that it's almost impossible to even tell who's winning most of the time!
As such, this system, no matter the power behind it, is doing something that dumb, brute-force, play-the-game AI written by world-experts in Go, AI, and game theory wasn't expected to be able to achieve within the next decade. And it primarily gets there because it learns from information fed to it.
For those who are more involved in AI research it is not so surprising. Similar general approaches to learning have been used in the "cognitive" branch of AI research for the last 15 years or so. The buzzword changed from "cognitive" to "deep learning" recently.
The key to success of AlphaGO is the position evaluation function that is learn from data. The surprise here is that learning from the game endings of internet GO players and somewhat informed computer vs computer games is enough to train an evaluation function with the predictive power to beat the world champion. In the old days of AI an expert-designed heuristic function would be used instead and a kind of smart position tree search would do the heavy lifting. But obviously this didn't work with GO due to combinatorial explosion and very difficult evaluation in the beginning and middle stages of the game.
I am not a mathematician, and I find this victory rather unimpressive and totally expected given the progress that have been made in machine learning in the last 20 years.
Go is rather simple compared to other problems like image recognition. The number of Go positions is dwarfed by the number of possible images (a 1M pixels color image leads to (3*255)^(10^6) possibilities - of course not all of them are valid and the manifold of relevant images is much smaller, but so is the manifold of relevant Go positions, I guess), and we've come up with pretty good results in those areas. Better than what humans can do in some areas.
The real question is: What is intuition? Is it something computable or not? If it is only some kind of statistical inference, then no wonder we are good at it: we have an inference engine which structure has been optimized by million years of evolution, and fed with bazillions of samples since our birth. But that doesn't mean it's impossible to build one as good as us. Sure the design of the model is trickier, but it's easier to feed the training samples given our technological ability to gather huge amounts of data.
I wonder when the term "true AI" will be ditched. To me, there is not "true AI" because there is no "AI" as opposed to "natural intelligence". The only difference is whether your computer is biological or electronic...
Video of some good progressive thrash music
It won't turn out that Se-Dol has quite a few other skills. That's the problem. There's too much focus on brainpower to solve these highly restricted set of problems. That's the issue. What makes real creativity is not a mind like Lee Se Dol (with respect). It's the people that are capable of inveting somethng truly original. AI can't do that .. yet. That's true creativity, and it's not something Lee Se dol has, or something that you find very easily in Asia, generally.
How would Einstein do against Lee Se Dol? Not ver y good..
Who would you put your money on to contribute to meaningful, original science? Einstein .. every time.
Even just this last week, another of Einstein's theories has been verified - gravitational waves, as shown by the LIGO istrument. He has more "creativity" than the sum totoal of all these Go players.
Playing a game, and doing it well, requires real creativity. Arguably a lot more than science, actually. When you study science, all you're doing is discovering information already out there - water had its properties and was built by molecules long before it was classified as H2O, and nothing changed after. Doing well at Go can't be calculated cold and hard - much of it is subjective, and that's what makes this discovery so important. The computer didn't win by just repeating the same patterns or evaluations over and over, but actually learned from each game and was able to apply that to the future. That's the start of self learning AI.
Like a ton of people in the world (the majority most likely), you apply the no true scotsman argument to this debate. It's not real AI until it learns strategies not programmed into it? Oh wait, no, it's not true AI until it creates its own strategies? Oh wait, no, it's not a true AI until it can do this to something other than Go? What next, it has to socialize and disobey? The approach this machine used was incredible, and the insight was extremely important - being able to learn by studying a history of decisions, that's something that lays the groundwork for every future AI project from here on out.
This represents a massive step forward in artificial intelligence, by leaps and bounds, and the sad part is, you don't even know it.
"Set a man a fire, he'll be warm for the rest of the night. Set a man afire, he'll be warm for the rest of his life."
It is not because you are completely uninterested by a subject that advancements in that subject have less "creativity" than in other subjects. I could reverse your 'demonstration' by saying that Einstein did get lucky that some of his results were proved true. He has been proved wrong about quantum entanglement. Should I compare him to Leonardo Da Vinci who was a genius painter, a great engineer and an anatomist and made great advance and publication in these subjects but was also interested in invention, sculpting, architecture, science, music, mathematics, literature, geology, astronomy, botany, writing, history, and cartography.
AlphaGo in its matches against Lee Sedol showed 3 interesting moves who will certainly be studied and played by all the professional Go players around the world for the next years. It is not because you limit yourself to a certain set of problems with the objective to excel at it that you are a lesser being that someone else.
Thank you for your response. Would you then agree that by your definition, a large majority of humans don't display and creativity?
See my journal for slashdot ID's by year. Mine created in 2005. http://slashdot.org/journal/289875/slashdot-ids-by-year
He's likely to be remembered as the last human being to beat a Go AI on tournaments.
Move 78, in particular, was so good that his partners and commentators in China have already called it "the hand of God", but it really was one of those things which happens once in a blue moon, even for a player like Sedol.
They actually got poker before Go, about a year or two ago, at least in heads-up situations.
The team who made alphago deserve credit, but their approach (from a high level) isn't so revolutionary. Go AI devs moved away from solely using brute force tree pruning (like what deep blue used) a long time ago.
The first big change was to use pattern recognition (matching sub-sections of the game with already known patterns) to prune faster. The second (and far more revolutionary) change was to apply an upper confidence bound based on a monte carlo simulation. This is where computers gained the ability to bypass those billions of moves with a margin for error. The third was the use of neural nets as a way to balance between brute force and pattern matching while managing the confidence levels of the monte carlo simulations.
The biggest difference with Alphago is corporate backing. I don't know how many people Google hired for the job, but the paper lists 20 (so probably more than that). Buying and running supercomputers is extremely expensive as well. With the exception of darkforest (Facebook's go machine which, as expected, appears to use a similar design), most teams consist of a very few people on small budgets without someone willing to spend millions to buy and run supercomputers for them.
Why would you want to look at all possible moves? Humans don't do this, why would it be necessary for computers?
Yes you do. You just use intuition to skip over moves that might not be worth your time, but you still consider them. AlphaGo does something similar with a neural network before brute-forcing into good possible moves.
Still, even if you don't want to consider 10^700 possible game trees on a clean Go board, the problem is still intractable. Go has, in average, 250 possible value moves to consider after each stone is placed. Chess has around 30.