CR4® - The Engineer's Place for News and Discussion®


The Engineer's Notebook Blog

The Engineer's Notebook

The Engineer's Notebook is a shared blog for entries that don't fit into a specific CR4 blog. Topics may range from grammar to physics and could be research or or an individual's thoughts - like you'd jot down in a well-used notebook.

Previous in Blog: Japan May Change Signage Ahead of 2020 Olympics   Next in Blog: Are Your Smarter (or Dumber) Than Your Ancestors?
Close
Close
Close
6 comments

AI Passes 'Go'

Posted March 21, 2016 10:52 AM by Hannes

Viewers of Darren Aronofsky's surreal film Pi in its entirety probably walked away with two lasting recollections: a man using a power drill to relieve his cluster headaches, and a romantic description of the ancient Chinese game Go. Unlike many other games of strategy, the Go board represents--as described by the character Sol Robeson--"an extremely complex and chaotic universe...and that is the truth of our world [...] it can't be easily summed up with math...there is no simple pattern."

Thanks to these attributes, Go has always been a major stumbling block for artificial intelligence. Late last year, however, Google DeepMind's AlphaGo program made a major breakthrough by defeating a highly ranked professional Go player sans handicap for the first time. More recently in a five-game match that took place from March 9-15, the program won four games against Lee Se-dol, one of the best players in the world. AI enthusiasts consider these matches major victories for the discipline, as many assumed that a competent Go program was still at least five years in the making.

Go's international appreciation is due to the fact that, despite having only two essential rules, there are a vast number of possible moves and unique games. Each player, one black and one white, takes turns placing their stones on the game board's intersections, typically forming chains. A chain which is completely surrounded by opposing stones is captured and removed from the board. A game may be scored by either the number of empty spaces a player's stones surround or the number of stones plus surrounded intersections. The only other rule is that a player may not make any move that returns the game to the position of the previous turn.

While the image on this page shows a 9x9 beginner board, professional games like those won by AlphaGo are played on a 19x19 board. AI expert Victor Allis estimates that a typical 19x19 expert game lasts 150 moves or so, with about 250 choices per move. These figures result in a game-tree complexity--a common measure of game complexity used in combinatorial game theory--of 10360. Compare this with tic-tac-toe's 26,830 or chess's 10120, and you start to see why Go is so difficult for the automated mind. (For those with enough time, patience, and interest, DeepMind's YouTube channel features all five March 2016 matches move-by-move.)

AlphaGo cracked the Go problem not by building a program adept at playing Go, like IBM's chess-playing Deep Blue, but by using a combination of more general machine learning and tree search algorithms to create a program adept at learning any game it practices and experiences. This step toward artificial general intelligence has met with differing reactions within the AI community. Most comment that it's probably a good time to discuss the social/cultural impact of this general intelligence; a year ago Stephen Hawking went so far as to suggest the possibility of a smart computer takeover. In response to the AlphaGo victories, Murray Campbell, an IBM scientist who worked on Deep Blue, more or less proclaimed the victorious end of AI board game experiments.

He might be right, but in my opinion there's no currently feasible leap from games to "real-life" general AI solutions, and computers obviously have a long way to go before (if ever) understanding human behavior beyond emulation. Carnegie Mellon's Claudico, a Texas hold 'em program that uses non-game-specific algorithms similar to AlphaGo's, lost a 2015 poker event against four top human players by over 700,000 chips. That program struggled with risky bets and bluffing, two behaviors difficult for a ruthlessly rational program to comprehend.

Much like Deep Blue in 1997 and Watson in 2011, AlphaGo's victory is one more AI milestone on the road to who knows where. But to quote Lee Se-dol after his historic defeat: "robots will never understand the beauty of the game the same way that we humans do."

Image credit: Jarrod Trainque / CC BY 2.0

Reply

Interested in this topic? By joining CR4 you can "subscribe" to
this discussion and receive notification when new comments are added.
Guru
Engineering Fields - Optical Engineering - Member Engineering Fields - Engineering Physics - Member Engineering Fields - Systems Engineering - Member

Join Date: Apr 2010
Location: Trantor
Posts: 5176
Good Answers: 616
#1

Re: AI Passes 'Go'

03/21/2016 1:11 PM

I have a copy of the movie Pi, and the soundtrack. I watch the movie at least once a year, and have a number of the tunes from the soundtrack on the memory stick in my car's audio system.
One of my all-time 20 favorite movies, and likewise one of my favorite music albums.

FWIW, fans of 'Breaking Bad' may recognize the actor (Mark Margolis) who played Sol Robeson, as the wheelchair-bound uncle (Tio Salamanca) of the character Tuco in 'Breaking Bad'. He is reprising the Tio role for the prequel series 'Better Call Saul'.

__________________
Whiskey, women -- and astrophysics. Because sometimes a problem can't be solved with just whiskey and women.
Reply
Guru

Join Date: Jun 2009
Location: South of Minot North Dakota
Posts: 8399
Good Answers: 770
#2
In reply to #1

Re: AI Passes 'Go'

03/21/2016 3:34 PM

Hmm. I will have to look into that movie and see what if it's something I would care for but to be honest pretty much every movie thatI have ever watched that had anything to do with mathematics or had a numerical value in its title was by my standards a total waste of my time.

23 with Jim Carry. WTF was that about?

Good Will hunting. Hey! A Robin Williams and Matt Damian movie! Cripes that was hard to sit through give their reputations as actors in my books.

A beautiful mind. A Russell Crow movie how could that go wrong. WTF? 2 hours and 15 minutes of watching a disturbed delusional math professor go on about things he largely imagined were real but weren't which by my estimation means that most of what he thought was mathematics was probably all imagined nonsense too.

From what I can see of PI the guy eventually drills a hole in his head which to be fair yea I can totally see where someone with mathematical obsessions needs that. I've been around a college level mathematic department staff enough times to get the very strong suspicions that most of them needed a fair amount of brain augmentations with a drill as well to get them in touch with reality.

Reply
Power-User
Hobbies - Musician - New Member Technical Fields - Technical Writing - New Member CR4 Admins - CR4 Moderator - New Member

Join Date: Nov 2012
Posts: 295
Good Answers: 4
#3
In reply to #2

Re: AI Passes 'Go'

03/21/2016 4:47 PM

I liked Pi--it's surreal and weird and deals with numerology and paranoia more than pure math. A Beautiful Mind inspired me to read the real bio of John Nash (didn't finish it), and I learned that he never suffered from visual hallucinations as depicted in the film--only auditory. That's pretty liberal cinematic license if you ask me.

__________________
"Freedom is nothing but a chance to be better." -Camus
Reply
Guru

Join Date: Jul 2012
Location: Madison, WI.
Posts: 1004
Good Answers: 42
#4

Re: AI Passes 'Go'

03/22/2016 10:11 AM

Go is an incredible game. it takes only minutes to learn and a lifetime to get a handle on. That a computer as finally been created that can work its way through the vast number of permutations per move in a reasonable amount of time is the only thing this actually demonstrates. After all, a computer does not produce strategy as such. It simply evaluates all possible moves and out comes and picks the one with the best numeric value.

In my mind this does not demonstrate an advancement in AI but an advancement in computational speed.

__________________
Knowing is the end result of learning, not believing.
Reply
Power-User
Hobbies - Musician - New Member Technical Fields - Technical Writing - New Member CR4 Admins - CR4 Moderator - New Member

Join Date: Nov 2012
Posts: 295
Good Answers: 4
#5
In reply to #4

Re: AI Passes 'Go'

03/23/2016 10:17 AM

I think the notability of recent AI breakthroughs like Watson and AlphaGo is that they combine very fast computation, as you mention, with higher degrees of machine learning so that mistakes are not repeated and the success of past moves dictates future moves. To me that's a pretty strategic component.

__________________
"Freedom is nothing but a chance to be better." -Camus
Reply
Guru

Join Date: Jul 2012
Location: Madison, WI.
Posts: 1004
Good Answers: 42
#6
In reply to #5

Re: AI Passes 'Go'

03/23/2016 11:04 AM

I do understand your suggestion. I still have to go back to the way computers work and look at your statement as simply a function of larger capacity memory and better addressing systems and faster processing and memory recall. A bad move would be given, for simplicities sake a negative operand so that when the processor is reviewing its memory for that particular move/situation it recalls the negative outcome and moves on to the next one.

What you suggest would simply be a larger memory, (vast) and a very good catalog and recall system functioning at high speeds. A pretty expensive option really. Instead I would posit that the system is functioning much more simply. Instead of carrying a vast library of previous games in memory I would go with the more direct and better suited to high speed computing approach and have it identify each potential placement, evaluate potential response and counter point to X number of moves out and select the one with the highest numerical probability of success. I would have to see the computer actually lay a trap for an opponent or push a bad position before I would consider that machine as "thinking". I believe that was what the small statement about poker was meant to illustrate. Machines still do not understand that a losing move, properly executed, can win because there is, as yet, no way to quanitfy the human response accurately.

Or so my thinking goes.

__________________
Knowing is the end result of learning, not believing.
Reply
Reply to Blog Entry 6 comments
Interested in this topic? By joining CR4 you can "subscribe" to
this discussion and receive notification when new comments are added.
Copy to Clipboard

Users who posted comments:

Hannes (2); rashavarek (2); tcmtech (1); Usbport (1)

Previous in Blog: Japan May Change Signage Ahead of 2020 Olympics   Next in Blog: Are Your Smarter (or Dumber) Than Your Ancestors?

Advertisement