AlphaGo Beats Top Player at Go - Share Your Thoughts

[Buzz Bloom]

Curious to see comments.

http://www.wired.com/2016/01/in-a-huge-breakthrough-googles-ai-beats-a-top-player-at-the-game-of-go/
http://phys.org/news/2016-01-chess-human-ancient-chinese-game.html​

Regards,
Buzz

 

[Greg Bernhardt]

Well there you have it, The Terminator is not far off.

 

[fluidistic]

Direct link to the paper: https://storage.googleapis.com/deepmind-data/assets/papers/deepmind-mastering-go.pdf .
What google did is a huge achievement, a leap of more than at least 10 years in go programming, leaving facebook's team and everyone else (see http://arxiv.org/abs/1511.06410) extremely far lagging behind.
Very original ideas that were never tried before, in alphago.
Granted Google bought some extremely talented programmers and also go players (e.g. Aja Huang). Marvelous work. An enormous step for humanity has been achieved.

 

[berkeman]

Can someone comment on why the game of Go is exponentially more complex than chess?

IN A MAJOR breakthrough for artificial intelligence, a computing system developed by Google researchers in Great Britain has beaten a top human player at the game of Go, the ancient Eastern contest of strategy and intuition that has bedeviled AI experts for decades.

Machines have topped the best humans at most games held up as measures of human intellect, including chess, Scrabble, Othello, even Jeopardy!. But with Go—a 2,500-year-old game that’s exponentially more complex than chess—human grandmasters have maintained an edge over even the most agile computing systems.

Sorry, I admit to not having read the full article. Thanks.

 

[jtdonova]

I believe one of the main reasons had to do with the number of possible moves in each game. Chess has on average about 20 moves per player where as Go has over 200. We can see how this gets exponentially more complex for a computer trying to determine the best move based on possible future moves. 20 options this turn, next turn, next turn 20^n for n moves. 200^n for n moves of Go.

 

[D H]

An enormous step for humanity sky net has been achieved.

Fixed that for you!

Can someone comment on why the game of Go is exponentially more complex than chess?

"There are more configurations of the board than there are atoms in the universe."

 

[Buzz Bloom]

Can someone comment on why the game of Go is exponentially more complex than chess?

Hi berkeman:

In addition to the factor of tree search multiplicity discussed in the posts by jtdonoval and D H, there is also the factor of pattern multiplicity. Using very rough and informal number estimates, a world class chess player may recognize perhaps 10,000 patterns which suggest plausible moves, while a world class go player may recognize 100,000.

Regards,
Buzz

 

[D H]

I played regularly in college, and then years later, in the mid 1980s. My boss at the time was 3 or 4 dan (amateur); we played almost every day after work. I worked my way up through the ranks to perhaps 2 dan (amateur). There were no nearby players when jobs made us part ways, so I bought one of the first computer go games available for sale to the public. This was Bruce Wilcox's Nemesis Go Master. I had to give it a nine stone handicap and then make three or four extremely dumb moves (i.e., right on the corners of the board) just to make it mildly interesting.That's akin to giving away your queen, both of your rooks, and then some in chess. My skills have since degraded due to lack of play and lack of a young mind; I'd be lucky to be 12 kyu now. Computer go on the other hand has improved vastly since then, but this latest development is a huge, huge leap.

 

[DiracPool]

If you think we've mastered artificial intelligence...



Better call Sol...

 

[Nantes]

Wow! I literally told my friend yesterday that no one has managed to make a Go bot that can beat a skilled human, even 20 years after Deep Blue beat world Chess champion Garry Kasparov, and then I randomly see this on the headlines here at PF :O

Elon Musk's warnings are starting to sound very real.

 

[Jeff Rosenbury]

Can someone comment on why the game of Go is exponentially more complex than chess?

In addition to the shear number of moves, there are many more strategic considerations in go than in chess.

Chess can be thought of in terms of territory covered, mobility of pieces, and tempo. (I'm sure there are more, but those are the big three.)

In go, there are considerations that don't even have English words. Sente is sort of like initiative, but also includes ideas of tempo and question asking. Other ideas include thickness, heaviness (not related to thickness; thick = good, heavy = bad), shape (how well your stones (pieces) work together), influence, territory, and elegance ("it just looks right"), tesuji (standard "trick" patterns), reading, life and death, and others. These all need to be weighed against each other each move (At least for humans, my understanding is skynet -- sorry -- AlphaGo uses a stochastic process where it fills the board with random moves, and keeps the ones that win -- as a part of its code anyway.)

All hail President Executron!

 

[Buzz Bloom]

Elon Musk's warnings are starting to sound very real.

Hi Nantes:

Since you did not show any emoticon, I am curious about your actual feelings regarding Musk's warnings. Do you find them to be actually scary, or were you being humorous, like Greg in post #2?

Regards,
Buzz

 

[jerromyjon]

Way to AlphaGo! I have a new game to learn I see... amazing I've never seen it, or noticed it if I did. Perhaps this can be a precursor to an AI rating system, Go level: alpha. The rest fall in the curve...

The Terminator is not far off

Come on already! I've only been dreaming about it forever... wish I was there.

 

[D H]

Way to AlphaGo! I have a new game to learn I see... amazing I've never seen it, or noticed it if I did.

There's only one site if you live outside of Asia and if you wish to learn the game. That site is Sensei's library, http://senseis.xmp.net . It is 100% free. There's only one option if you live outside of Asia and you wish to play the game, and that's to install an IGS client on your computer and then play against Asians at odd hours of the day (odd hours to you, that is). You can find plenty of IGS clients. Sensei's Library has a large catalog.

I'm over 60; I can't play at IGS anymore. Go is a game for those with extremely agile, extremely intuitive, and extremely imaginative minds. In this regard, go is a bit like chess. Go is a game best suited for younger adults. The world champion against whom AlphaGo will play next is 33. That's a bit old for a world champion.

Come on already! I've only been dreaming about it forever... wish I was there.

Be very, very careful of what you wish for. You might well get what you wish for.

DeepMind, the Google subsidiary that created AlphaGo, has created an AI ethics board. What this means, who knows? Playing go at anything beyond the potzer stage requires imagination. (Full disclosure: I gave up chess for go 40+ years ago. Playing chess is mechanical. Playing go is anything but.)

An algorithmic model of "imagination" is exactly what DeepMind claims to have accomplished. This leaves me split in two. One part of me says "This is so cool!". Another part says SKYNET! (Oh noes!)

 

[jerromyjon]

The world champion against whom AlphaGo will play next is 33. That's a bit old for a world champion.

I'm 44 but I've never "stretched" those mental muscles yet, and imagination is by far my best quality. And while I'm learning the game I'll be thinking about how to "spot" patterns to exploit which could take mastery then perhaps I can join the fun and create my version of a quantum supercomputer brain in my garage.

 

[D H]

More full disclosure: I took on a new job 35+ years ago. My previous job made me feel a need to cleanse me soul. (Anything more than that, deponent saith not.) My new job sent me all over the world to help develop and install meteorological ground stations, and that included Communist China. This was well before Tiananmen Square. We were assigned "watchers" by the Chinese government, and we knew full well who they were. When my watcher asked me what I most wanted to buy before I went home, I responded that what I wanted most was a pair of sets of jade go pieces and go board to go along with it. We went to a low-scale shopping center, and there it was.

I still have that set. The stones make the most wondrous sound when one intentionally places one of them on a point.

 

[jedishrfu]

The iOS universe has the the SmartGo Kifu app that is pretty decent. It has features to teach the game as well as playing it. It's a bit pricey as compared to other iOS apps at around $20 but it has a cheaper cousin SmartGo Player at $3 with more limited features.

https://www.smartgo.com/kifu.html

It also has Windows and MacOS versions as well but no Android version that I can see.

The cheaper cosin SmartGo Player features vs SmartGo Kifu:

https://www.smartgo.com/player.html#compare

 

[atyy]

DeepMind, the Google subsidiary that created AlphaGo, has created an AI ethics board. What this means, who knows? Playing go at anything beyond the potzer stage requires imagination. (Full disclosure: I gave up chess for go 40+ years ago. Playing chess is mechanical. Playing go is anything but.)

An algorithmic model of "imagination" is exactly what DeepMind claims to have accomplished. This leaves me split in two. One part of me says "This is so cool!". Another part says SKYNET! (Oh noes!)

The news reports about AlphaGo quote Ilya Sutskever of OpenAI.

https://openai.com/blog/introducing-openai/

Aren't Asimov's 3 laws enough?

 

[Buzz Bloom]

I hope no one minds if I make a prediction regarding the near future of go AIs performance vs. human opponents. My prediction is based on my recollections of the history of chess AIs when they first became strong enough to being to win against strong (as I recall master level) human opponents. After less than a year, the chess AIs good performances began to decline significantly. I recall that the reason for the early success and later decline is that the inherent stylistic weakness of those AIs were not at first recognized, but after a while human opponents figured it out. The weakness was that the AIs were excellent at tactics, but had no programmed concepts related to playing positional chess. Later AI generations improved by a combination of deeper analysis and including some aspects of positional play in the position evaluation algorithms.

I predict that the go AIs, like AlphaGo, will go through a similar performance experience over the next year or so. And when that happens, the developers will then find a way to make another major advance.

BTW, I also believe that chess AIs, as good as they have become, have the possibility of further major improvement that would make them virtually unbeatable by humans. The improvement will involve adding the ability to set deep traps. If there is interest, I will explain this idea in more detail in a separate thread.

 

[Khashishi]

I thought chess AIs are already unbeatable by humans. You mean MORE unbeatable?

 

[CrackerMcGinger]

How can I get a computer version of Go? I'm a pretty good chess player, but frustrated at the lack of options I can make for moves.

 

[Buzz Bloom]

I thought chess AIs are already unbeatable by humans. You mean MORE unbeatable?

Hi Khashishi:

In recent years Chess AIs have won matches vs very strong players, including matches with odds being given to the human, so you are mostly right.

https://en.wikipedia.org/wiki/Human–computer_chess_matches​

However, the AIs are not quite yet totally dominant, since the humans win sometimes. Also, many of the AI wins have occurred after a very bad move being made by the strong human player. A reasonable interpretation might be that the humans are psychologically not adequately prepared for these contests.

Regards,
Buzz

 

[phinds]

Aren't Asimov's 3 laws enough?

I assume this was tongue in cheek but in any case it's really unfortunate that Asimov's 3 laws are a total joke in practical terms. Our future would likely be much safer if that were no the case.

 

[PAllen]

Hi Khashishi:

In recent years Chess AIs have won matches vs very strong players, including matches with odds being given to the human, so you are mostly right.

https://en.wikipedia.org/wiki/Human–computer_chess_matches​

However, the AIs are not quite yet totally dominant, since the humans win sometimes. Also, many of the AI wins have occurred after a very bad move being made by the strong human player. A reasonable interpretation might be that the humans are psychologically not adequately prepared for these contests.

Regards,
Buzz

I don't think it is so much psychology as consistency. Using a top engine to analyze even the best ever games between humans show exploitable errors made by the stronger player. On the other hand 'centaurs' consisting of strong human player plus a strong but not top engine, still consistently win matches against the top engines. Thus, with the computer assist to avoid tactical oversights, the human still wins.

A criteria for when an computer is better in all ways than top human players would be when a centaur with slightly weaker engine loses e.g. a 10 game match to the stronger engine, or only draws at best (by doing nothing) if the the engines are equally strong.

 

[Hornbein]

Here is a little article I wrote about it. It's too late at night to customize if for this group, so here it is.

Computers have traditionally played games by exhaustively searching all possible moves. Nobody thinks that this is real intelligence.

Mastering the game of Go has long been seen as the benchmark of true artificial intelligence because it can't possibly be done by brute force search. The computer has to think like a human Go master. In October 2015 a computer program defeated the Go champion of Europe five games to zero. It did it by pure intuition, searching no moves at all! Artificial intelligence is here.

http://googleresearch.blogspot.com/2016/01/alphago-mastering-ancient-game-of-go.html

It's been in the works since the late fifties. The basic idea of how to do it was invented by Marvin Minsky, a workmate of Noam Chomsky at MIT. After nearly sixty years the project has finally borne fruit, financed by a speck of the one hundred billion dollars in capital controlled by Google.

The computer has to learn subtle inexplicable patterns. In other words, the computer has to develop intuition. The machine acquired this by playing untold millions of games against other programs and against itself.

Even more impressive to me is that a computer learned to play dozens of simple Atari computer games by looking at raw pixel inputs. That is, they were given seemingly senseless sequences of numbers as input. They were given an equally senseless number of actions that they could perform as output. Finally they were given a score as to the results of such actions. With no knowledge of the rules of the game whatsoever, using pure intuition the computer was able to learn to play the games better than a human.

This is called inductive reasoning. Computers have long been better than humans at deductive if-then reasoning. Now they are better at inductive reasoning too. What's left? The only remaining advantage people have is that they are able to learn from fewer experiences. Humans don't need to play millions of games to find a pattern. But how long will this difference last? Not long, I think.

According to The Atlantic, pattern-learning programs like this are already in use to make job hiring decisions. The online records of candidates are fed into a computer and it gives a hire/not hire score.

Already a system is in place for humans to provide input into machine learning networks. It's run by Amazon and is called The Mechanical Turk. People log in, perform simple tasks, and are paid something like one dollar an hour. You may sign up right now if you like.

Surely the irony in this -- well, it's like something out of science fiction. The Mechanical Turk was a fake chess playing automaton. Gears whirled around, but it was actually powered by a dwarf hidden inside. Nowadays the roles are reversed. Instead of a computer using a person to cheat, people use may computers in order to cheat at games.

A Reproduction of the Mechanical Turk. Sorry, I don't know how to change the size.

So: computers are smarter than people. They also have access to a lot more data than does a person. Where will this all lead? With Amazon's Mechanical Turk service, we now have people performing unskilled labor for computers and return being paid wages. Very low wages.

Oh well. I hope it all turns out for the best. A match with the Go champion of the world is scheduled for March in Seoul, South Korea. Welcome to Minskyworld.

 

[DiracPool]

Computers have long been better than humans at deductive if-then reasoning. Now they are better at inductive reasoning too. What's left? The only remaining advantage people have is that they are able to learn from fewer experiences.

Let's not get carried away here. AlphaGo has done some interesting things, and this is all great. My issue is that these systems are largely what their moniker indicates, "artificial intelligence." And this is not to be confused with artificial as in a substitute for actual human intelligence. It means artificial in that it has little relationship to how the brain actually produces intelligent thought even though the term "deep neural networks" and other putative synonyms are frequently bandied about in these discourses.

Although AlphaGo may employ some clever heuristics that go beyond traditional "tree search" methods, it is more or less a "one-trick pony," as was Deep Blue for chess and Watson for Jeopardy. What characterizes human intelligence is as much its generality as it is it's deep cognitive structure. It's the ability for it to play a "go" game at 5PM, cook a crab bisque at 8PM, write a pop song at 10PM, and make whoopie with the spouse at 11:30. And then waking up at 6AM and driving to work.

So to say that computers are smarter than humans is to really really miss what human intelligence is. It is about intuition but it's more about the capacity for creativity (in art and in science) than it is about some formalized notion of what intuition is per se. I guess it just irritates me a bit when people throw around the concept of human intelligence as if it were something we could isolate and quantify with a measure we can compare to a computer algorithm. Even the OP's posted article doesn't claim this despite it's sensational title:

http://www.wired.com/2016/01/in-a-huge-breakthrough-googles-ai-beats-a-top-player-at-the-game-of-go/

But DeepMind’s system is very much under the control of Hassabis and his researchers. And though they used it to crack a remarkably complex game, it is still just a game. Indeed, AlphaGo is a long way from real human intelligence—much less superintelligence. “This is a highly structured situation,” says Ryan Calo, an AI-focused law professor and the founder of the Tech Policy Lab at the University of Washington. “It’s not really human-level understanding.”

I've been involved in the AI field since the late 80's and have seen many promising technologies come and go. You name it, I've seen it, PDP, "fuzzy logic," simulated annealing, attractor neural networks, reinforcement learning etc. etc. Each one of them promised the same as the article stated above...

But it points in the direction. If DeepMind’s AI can understand Go, then maybe it can understand a whole lot more. “What if the universe,” Calo says, “is just a giant game of Go?”

Do you know how many times I have heard this? In 1996 I actually worked with one of the principal research teams that were developing the technology of optical holographic memory in crystals. Do you remember this? I was extremely excited as to where this technology might take us. At the time, Star Trek Voyager was on the tele and I remember sitting back watching it on my trusty old CRT thinking we were a breath away from making Star Trek a reality with this new technology. Good times. AI, here we come. That was in 1996. It didn't happen. So, I'm not trying to discourage optimism here, far from it, I'm just saying that equating the mechanism that makes Alpha Go with the mechanism that generates human cognition is like comparing apples to oranges, and making a statement that AlphaGo is "smarter" than humans is just simply naive.

Welcome to Minskyworld.

Don't count on it. If Minsky had the answer we would see some actual human-like intelligent robots walking around and we don't. He had half a century to make this happen and what happened? Nothing.

 

[Monsterboy]

I have never played Go but i have played a lot of chess ,this is a new must learn game for me now ,i am very interested to learn the difference in the kind of thinking involved ,any advise from someone who knows both the games will be appreciated.

 

[CrackerMcGinger]

A Reproduction of the Mechanical Turk. Sorry, I don't know how to change the size.

Oh my... that is huge...

 

[Buzz Bloom]

Hi @Monsterboy:

I suggest three avenues for learning the game of go.

1. Find an experienced player who is willing to play with you. If s/he is also willing to explain aspects of the game as you play, or after a game, so much the better.
It is likely that any experienced player can play against you with a large handicap and still win. One of the benefits of go, as compared with other games like chess, is that it is simple for the stronger player to play with a handicap (large or small) against a weaker player without seriously distorting the nature of the game. As a chess player you are no doubt experienced with keeping a record of the game as you play. It is just as valuable to do that in go so you can play over your games and learn from mistakes.

2. If your experienced opponent is willing, begin learning the game playing on a 13x13 board rather than the standard 19x19 board. Most of the aspects of the game you want to gain experience with as early as possible are present with the smaller board. Also the games much less time to play. A 5 stone handicap on the 13x13 board is about the equivalent of a 9 stone handicap of the 19x19 board.

3.The are a very large number of excellent books about go, although perhaps not as many as there are chess books. The book I recommend as a first book is:

Life and Death by James Davies (Elementary Go Series, Vol. 4) The Ishi Press (1975).​

The is about the heart of local tactics.

Regards,
Buzz

 

[Planobilly]

Can someone define the meaning of the word "intuition" as it is used in this context?

This is a quote from one of the links.

These networks don’t operate by brute force or handcrafted rules. They analyze large amounts of data in an effort to “learn” a particular task. Feed enough photos of a wombat into a neural net, and it can learn to identify a wombat. Feed it enough spoken words, and it can learn to recognize what you say. Feed it enough Go moves, and it can learn to play Go.

How can this be defined as "intuition"?

I think AI will prove to be very useful and can solve some very pressing issues. Go appears to be a good test platform for the development of AI.

Cheers,

Billy

 

[jerromyjon]

I assume this was tongue in cheek but in any case it's really unfortunate that Asimov's 3 laws are a total joke in practical terms.

I don't think that is so far off from the basic concept, I've given this ALOT of thought. I'm not saying it is as simple as typing 3 strings of characters and sticking them somewhere in memory and they can't hurt us it is a hierarchy of laws to keep in mind while building robots, and again it's not as simple as "everyone smart enough to make robots will be smart enough to make them safe for human interaction" but it is like the foundation of religion for robot designers. I don't even agree with the first law simply stating "human" as if any other forms of life are of less importance.

 

[phinds]

I don't think that is so far off from the basic concept, I've given this ALOT of thought. I'm not saying it is as simple as typing 3 strings of characters and sticking them somewhere in memory and they can't hurt us it is a hierarchy of laws to keep in mind while building robots, and again it's not as simple as "everyone smart enough to make robots will be smart enough to make them safe for human interaction" but it is like the foundation of religion for robot designers. I don't even agree with the first law simply stating "human" as if any other forms of life are of less importance.

Well, we're going to have to agree to disagree on this. I seriously think they are a joke and I've thought about them since reading Asimov in the 50's. Loved the stories despite their implausibility.

 

[Buzz Bloom]

How can this be defined as "intuition"?

Hi @Planobilly:

I am not sure why you ask this, but I assume you have some concept about "intuition" that is distinctly different than

These networks don’t operate by brute force or handcrafted rules. They analyze large amounts of data in an effort to “learn” a particular task. Feed enough photos of a wombat into a neural net, and it can learn to identify a wombat. Feed it enough spoken words, and it can learn to recognize what you say. Feed it enough Go moves, and it can learn to play Go.

Since I have no intuition about your concept of "intuition", I will start with some definitions from the Internet.
From https://www.wordnik.com/words/intuition

from The American Heritage® Dictionary of the English Language, 4th Edition

• n. The act or faculty of knowing or sensing without the use of rational processes; immediate cognition. See Synonyms at reason.
• n. Knowledge gained by the use of this faculty; a perceptive insight.
• n. A sense of something not evident or deducible; an impression.

These definitions seems to emphasize that intuition is a process for acquiring understanding/knowledge that is not "rational", where "rational" implies what has been metaphorically called the "left-brain" functions. That is, intuition is metaphorically a right-brain function. Another distinction might be, rationality is a step-by-step sequential rule-based deductive process while intuition is a gestalt holistic process. The quote from your post is making that distinction. That is, intuition is a non-rational response to a given situation based on an accumulation of previous experiences which are not consciously remembered when the intuitive response occurs.

I hope this is helpful.

Regards,
Buzz

 

[D H]

I have never played Go but i have played a lot of chess ,this is a new must learn game for me now ,i am very interested to learn the difference in the kind of thinking involved ,any advise from someone who knows both the games will be appreciated.

I'll give two additional alternatives to @buzz Bloom's already excellent answer. One is the 80+ year old book by Edward Lasker, Go and Go-Moku. It's old, it's outdated, but it's cheap and it is written by someone fluent in English. Edward Lasker was a chess grandmaster who later found go to be a superior game. (And it is.)

The other alternative is the internet. There are lots of possibilities here, but the starting point has to be (IMHO) Sensei's Library, http://senseis.xmp.net .

 

[Planobilly]

Hi Buzz,

Thanks for responding. Human intelligence and artificial intelligence are two vastly different things. Human intuition and computer intuition are vastly different things if computer intuition can even be said to exist in the first place.

Based on The American Heritage® Dictionary of the English Language, 4th Edition definition of the word intuition is the AI computer program doing any of the following?
1. knowing or sensing without the use of rational processes?
2. gaining knowledge by a non rational process?
3. is the AI computer program sensing something not evident or deducible?

I don't have my own special definition of the meaning of intuition. I just don't think one can call this AI program intuitive based on the standard definition of the word.
We like to imbue objects both animate and inanimate with human characteristics which makes for interesting cartoons. Not so much in serious discussions of technology.

The above is not stated to detract from the considerable value of the work done by google and the advancement in AI that the work represents. The AI program stands on it's own merit and has no need to be embellished by comparison to ill defined terms or conditions.

Cheers,

Billy