ISF Logo   IS Forum
Forum Index Register Members List Events Mark Forums Read Help

Go Back   International Skeptics Forum » General Topics » Science, Mathematics, Medicine, and Technology
 


Welcome to the International Skeptics Forum, where we discuss skepticism, critical thinking, the paranormal and science in a friendly but lively way. You are currently viewing the forum as a guest, which means you are missing out on discussing matters that are of interest to you. Please consider registering so you can gain full use of the forum features and interact with other Members. Registration is simple, fast and free! Click here to register today.
Reply
Old 12th January 2018, 11:46 AM   #81
3point14
Pi
 
3point14's Avatar
 
Join Date: Nov 2005
Posts: 15,851
Originally Posted by Meadmaker View Post
Sort of. It didn't get fed "the correct answers", it just tried things to see if it ended up with "the correct answer", i.e. a victory.

And it's hard to say exactly what it figured out. One thing about neural nets is that they are pretty opaque. No one really knows what "rules" it was using, or if they do they haven't published it yet.


For what it's worth, one of the things that made its style of chess so "alien", as it was described in the press conference, was that it seemed incredibly willing to sacrifice material to gain a positional advantage. I'm sure it didn't trade queens for pawns all that often, but it would be more likely to do that than a human player would.

As I understand it, it also showed a willingness to repeatedly move the same piece which is something human players are reluctant/taught not to do.
__________________
Up the River!
3point14 is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 11:49 AM   #82
Fast Eddie B
Philosopher
 
Fast Eddie B's Avatar
 
Join Date: Sep 2010
Location: Mineral Bluff, GA
Posts: 5,274
Originally Posted by Meadmaker View Post
For what it's worth, one of the things that made its style of chess so "alien", as it was described in the press conference, was that it seemed incredibly willing to sacrifice material to gain a positional advantage.
But human players do that as well.

I think what makes it “alien”, is that all humans start out playing chess by learning standard values for the pieces, set openings and common patterns and tactics. It kind of forces them into a form of “groupthink” that this program is immune from, allowing it to find the best moves without being saddled with preconceived notions.

That said, I wonder how easy it is to tell a computer opponent from a human one, in a Turing Test sort of way. The best move is the best move, whether found by a human brain or a silicon one.
__________________
“I do solemnly swear (or affirm) that...I will to the best of my ability, preserve, protect and defend the Constitution of the United States.” - President Donald J. Trump, January 20, 2017.
"And it's, frankly, disgusting the way the press is able to write whatever they want to write. And people should look into it." - President Donald J. Trump, October 11, 2017.
Fast Eddie B is online now   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 11:57 AM   #83
sir drinks-a-lot
Illuminator
 
sir drinks-a-lot's Avatar
 
Join Date: May 2004
Location: Cole Valley, CA
Posts: 3,419
Cool thread.

To get an idea of the kinds of things that will probably pose some of the greatest challenges on the road to AGI, consider capthcas. These are the little dialogs when you register for a website that show you an image and ask you to type the characters you see. I think the ones that ask you context-sensitive questions like "how many vehicles are in this photo" are even better.

Captchas work so well because the human mind is so great at understanding "blurry" concepts like the letters of the alphabet. Take a look at all of the fonts on your computer, or even better yet, some of the more unusual fonts that are available to install. What, exactly, constitutes a valid capital A? It is, I think, not possible to define a set of criteria that must be met by a written or printed character to discern whether or not it is a capital A. Try it! Almost any rule can be violated while still being left with something humans will see as an "A". The same goes for the rest of the alphabet.

Additionally, if people are allowed to see a whole word the characters can get more and more abstract. For example, if the "A" is rendered in a certain unusual way (for example, split in two along the vertical axis), one might guess that an "H" in the same font would be constructed in a similar way. And, even further than this, the whole word as a whole can be messed with. For examples here, one only need to look at the logos of popular rock bands such as Yes or Aerosmith. And don't even get started with Metal bands!
__________________
I drink to the general joy o' th' whole table. --William Shakespeare

Last edited by sir drinks-a-lot; 12th January 2018 at 11:59 AM.
sir drinks-a-lot is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 12:04 PM   #84
Meadmaker
Penultimate Amazing
 
Meadmaker's Avatar
 
Join Date: Apr 2004
Posts: 18,317
Originally Posted by Fast Eddie B View Post
But human players do that as well.

I think what makes it “alien”, is that all humans start out playing chess by learning standard values for the pieces, set openings and common patterns and tactics. It kind of forces them into a form of “groupthink” that this program is immune from, allowing it to find the best moves without being saddled with preconceived notions.

That said, I wonder how easy it is to tell a computer opponent from a human one, in a Turing Test sort of way. The best move is the best move, whether found by a human brain or a silicon one.
My rating never got above 1000, so I'm not exactly a chess expert.

However, from what I have read, alphazero was a lot more likely than humans to sacrifice material for position. At least, that's what humans figured it must be doing. No one could ask alphazero. They just saw, from what I have read, that it won a lot of games when playing "behind" in material.

As for the second paragraph, I think you are right. It will be interesting to see if any humans can successfully imitate alphazero's style.
Meadmaker is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 12:16 PM   #85
sir drinks-a-lot
Illuminator
 
sir drinks-a-lot's Avatar
 
Join Date: May 2004
Location: Cole Valley, CA
Posts: 3,419
Originally Posted by Fast Eddie B View Post
But human players do that as well.

I think what makes it “alien”, is that all humans start out playing chess by learning standard values for the pieces, set openings and common patterns and tactics. It kind of forces them into a form of “groupthink” that this program is immune from, allowing it to find the best moves without being saddled with preconceived notions.
Good point. When I play, I'll notice that my opponent or I seem to be using one of the standard specific named openings or variations, like the Ruy Lopez or something similar. The understanding among most players that the reason these are so common is that they are the best moves and responses given how the game got off.

I'm not a very good player by any stretch, but about 15 years ago I had a roommate who was really into chess, and one of his friends was one of the best in the country at the time. I was on the periphery of some great games and chess discussion. Then my roommate was going through famous games and discovered a particular game where a player opened with what looked like the worst possible opening move* and ended up exploiting what little advantage there was after the move and winning the game. My roommate was so impressed that he started developing and using that same opening move to throw opponents off. He would win the games more often than one would expect.

I'd imagine chess-playing computers could have a similar effect.

* IIRC, the move was, while playing white, to advance the rook pawn a single square. I forget which of the rook pawns. This looks like a horrible move as it secures less space than moving two squares, doesn't really seem to do much to secure the strategic center of the board, and doesn't really open up any other pieces for motion. I can only think of a few other moves that appear to be weaker openings; moving one of the rook knights in front of it's rook pawn or, of course, resigning.
__________________
I drink to the general joy o' th' whole table. --William Shakespeare

Last edited by sir drinks-a-lot; 12th January 2018 at 12:19 PM.
sir drinks-a-lot is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 01:22 PM   #86
3point14
Pi
 
3point14's Avatar
 
Join Date: Nov 2005
Posts: 15,851
Originally Posted by sir drinks-a-lot View Post
* IIRC, the move was, while playing white, to advance the rook pawn a single square. I forget which of the rook pawns. This looks like a horrible move as it secures less space than moving two squares, doesn't really seem to do much to secure the strategic center of the board, and doesn't really open up any other pieces for motion. I can only think of a few other moves that appear to be weaker openings; moving one of the rook knights in front of it's rook pawn or, of course, resigning.
I'm aware there's a line that starts a4. I don't think I've ever seen a3.


Found it:

https://en.wikipedia.org/wiki/Ware_Opening
__________________
Up the River!
3point14 is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 01:54 PM   #87
ginjawarrior
Thinker
 
Join Date: Feb 2011
Posts: 214
Originally Posted by 3point14 View Post
I'm aware there's a line that starts a4. I don't think I've ever seen a3.


Found it:

https://en.wikipedia.org/wiki/Ware_Opening
you have the clemenz or the Anderson openings

https://en.wikipedia.org/wiki/Clemenz_Opening

https://en.wikipedia.org/wiki/Anderssen%27s_Opening

YouTube Video This video is not hosted by the ISF. The ISF can not be held responsible for the suitability or legality of this material. By clicking the link below you agree to view content from an external website.
I AGREE


don't think ive ever seen them
ginjawarrior is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 02:16 PM   #88
Beelzebuddy
Philosopher
 
Beelzebuddy's Avatar
 
Join Date: Jun 2010
Posts: 6,273
Originally Posted by sir drinks-a-lot View Post
To get an idea of the kinds of things that will probably pose some of the greatest challenges on the road to AGI, consider capthcas. These are the little dialogs when you register for a website that show you an image and ask you to type the characters you see. I think the ones that ask you context-sensitive questions like "how many vehicles are in this photo" are even better.
Fyi, google's current captchas, "click the squares with vehicles" and what not, are literally providing training data for their automated driving algorithms.
Beelzebuddy is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 02:32 PM   #89
Thor 2
Illuminator
 
Thor 2's Avatar
 
Join Date: May 2016
Location: Brisbane, Aust.
Posts: 3,976
Originally Posted by Fast Eddie B View Post
But human players do that as well.

I think what makes it “alien”, is that all humans start out playing chess by learning standard values for the pieces, set openings and common patterns and tactics. It kind of forces them into a form of “groupthink” that this program is immune from, allowing it to find the best moves without being saddled with preconceived notions.

That said, I wonder how easy it is to tell a computer opponent from a human one, in a Turing Test sort of way. The best move is the best move, whether found by a human brain or a silicon one.

Yes human players who really know the game certainly do.

I can play the game to a reasonable standard, and find when I play someone who is new to the game, they often show reluctance to exchange pieces. Chess is a ruthless game, and you should be prepared to slaughter and be slaughtered to achieve checkmate.

Some years ago I met a guy who said he played chess but followed this statement with the proviso that ... "I never play across the board. I only play online."

Somewhat surprised I agreed to play him online and won the first 3 games, but then things changed and he started to beat me! His style of play seemed to have changed as well and I suspect I was not playing him, but a computer program he was using.

Interestingly I found the experience of playing online not so enjoyable. You were allowed a day or more to make a move, and I found it time consuming and stressful, because of the number of moves ahead you could analyse, in the time allowed.
__________________
Thinking is a faith hazard.
Thor 2 is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 02:42 PM   #90
sir drinks-a-lot
Illuminator
 
sir drinks-a-lot's Avatar
 
Join Date: May 2004
Location: Cole Valley, CA
Posts: 3,419
Originally Posted by Thor 2 View Post
Yes human players who really know the game certainly do.

I can play the game to a reasonable standard, and find when I play someone who is new to the game, they often show reluctance to exchange pieces. Chess is a ruthless game, and you should be prepared to slaughter and be slaughtered to achieve checkmate.
Yeah. I'm known for swapping queens if I think it's to my advantage, and a lot of players who are new to the game get super annoyed at this for whatever reason.
__________________
I drink to the general joy o' th' whole table. --William Shakespeare
sir drinks-a-lot is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 03:27 PM   #91
Fast Eddie B
Philosopher
 
Fast Eddie B's Avatar
 
Join Date: Sep 2010
Location: Mineral Bluff, GA
Posts: 5,274
There's been a long-simmering debate as to the best first move for white.

The two most common are moving the King pawn or Queen pawn ahead 2 squares (e4 or d4*). Both have their defenders and detractors.

Seems like AlphaZero may have settled that debate - sort of.

Of the 10 games released, AlphaZero played white in 8.

6 of those began with d4. The remaining two were Nf3, but transposed quickly into standard QP fare.

A link to the games released: http://www.chessgames.com/perl/chess...944&pid=160016

Hope they plan or publishing the rest of the games for a more complete picture.


* P-K4 or P-Q4 for us old-timers.
__________________
“I do solemnly swear (or affirm) that...I will to the best of my ability, preserve, protect and defend the Constitution of the United States.” - President Donald J. Trump, January 20, 2017.
"And it's, frankly, disgusting the way the press is able to write whatever they want to write. And people should look into it." - President Donald J. Trump, October 11, 2017.

Last edited by Fast Eddie B; 12th January 2018 at 03:28 PM.
Fast Eddie B is online now   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 12th January 2018, 06:01 PM   #92
lpetrich
Muse
 
Join Date: Feb 2007
Posts: 755
From the looks of it, AlphaZero is using a very atypical strategy. Could that be contributing to its success?

It would be interesting to see how well it does with chess variants like Crazyhouse, King of the Hill, Three-Check, Chess-960, and Losers/Suicide.
lpetrich is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 13th January 2018, 06:00 AM   #93
Meadmaker
Penultimate Amazing
 
Meadmaker's Avatar
 
Join Date: Apr 2004
Posts: 18,317
Originally Posted by lpetrich View Post
From the looks of it, AlphaZero is using a very atypical strategy. Could that be contributing to its success?

It would be interesting to see how well it does with chess variants like Crazyhouse, King of the Hill, Three-Check, Chess-960, and Losers/Suicide.
I would guess that it would do amazingly well with all of those, with the possible exception of Chess960. Thinking about why reveals something about the difference between humans and alphazero.

All of those games are just abstract strategy, perfect information games. Just like chess, go, and shogi, which alphazero has mastered, all information is known. Players move alternately. There's no luck. Someone would program the board representation and legal moves, and then it would teach itself to play those games.

The thing that is different between humans and alphazero is that if a human expert is presented with a new variant, he will be able to play that new variant pretty well, because the human plays using principles of chess. As long as those principles can be applied in the new game, a human chess master will beat a human chess novice.

For alphazero, each game would be pretty much brand new, and it would just train itself to play the new game. It only takes a few hours to play enough games to become world champion.

The thing that would make chess960 a possible exception is that I don't know how much difference there would be in the 960 possible variants. Would it have to treat each one as 960 different games? There is plenty of overlap in play, of course, so I am guessing alphazero would do extremely well. With its apparent dominance of position-oriented chess, it ought to do well, but perhaps it would encounter so many never before seen positions that it would fail, or require a great deal more learning time.

I can think of two variants that it might be difficult for alphazero to deal with. In Bughouse, there is a time element. Sometimes, it is best for one player to just sit and do nothing, but you have to be very careful about how you manage your clock in those cases. You have to play differently when you may run out of clock time than when you have lots of time available. The amount of time available on each of the four players' clocks can dictate the best move. "I have more time than my opponent", is something harder to represent and pass as an input to a neural network.

The other type of variant that might be difficult are the various large board/ high piece count variants. I've played Tamerlane's Chess. It's not that hard. It has 11 different types of pieces, 28 initial pieces per player, and an 11x10 board. (And just to make it fun, each of the 11 pawns is actually slightly different. Would alphazero treat it as 21 different pieces?) I am sure alphazero could master it, but it would be interesting to see how long it would take to train, or how much hardware is required. However, now jump to something ridiculous, like Taikyoku Shogi. https://en.wikipedia.org/wiki/Taikyoku_shogi

I suspect that the number of pieces/size of board would be seen something with an exponential complexity to alphazero, whereas for a human shogi expert, Taikyoku Shogi would just be a ridiculously large variant. A human would just break it down into some local areas, and muddle through.

That's where the creators of alphazero will have to be able to tell us whether they've created a true breakthrough in AI, or just an interesting novelty, possibly to be incorporated later into something that might be more worthy of the term "intelligence".

Last edited by Meadmaker; 13th January 2018 at 06:02 AM.
Meadmaker is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 14th January 2018, 02:42 PM   #94
Thor 2
Illuminator
 
Thor 2's Avatar
 
Join Date: May 2016
Location: Brisbane, Aust.
Posts: 3,976
Interesting to observe the tussles between these different programs and I ponder about why when considering what Fast Eddie B wrote.

Quote:
Of the 10 games released, AlphaZero played white in 8.

6 of those began with d4.
I am wondering why AlphaZero did not play d4 in all of the games? Is it just that learning is still happening or is the machine playing a tactic to try and unsettle the opponent. The latter would indicate some deeper cognitive ability perhaps?

Be interesting if we could sense and gauge reactions in the circuitry. An excited pulse here or there, when a telling move is played, or a frantic flurry when danger is sensed. I suggest if we become aware that the machine cares about winning or losing then we may be in trouble.
__________________
Thinking is a faith hazard.
Thor 2 is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 14th January 2018, 03:43 PM   #95
Fast Eddie B
Philosopher
 
Fast Eddie B's Avatar
 
Join Date: Sep 2010
Location: Mineral Bluff, GA
Posts: 5,274
Originally Posted by Thor 2 View Post
I am wondering why AlphaZero did not play d4 in all of the games?
I wondered that as well.

Two options I see...

1) During the course of the match it gained a new insight into what the best first move was, or...

2) After playing myriad games with itself, it found that even against itself it won more often if it was not 100% predictable in its first move.

Still, a fascinating question.
__________________
“I do solemnly swear (or affirm) that...I will to the best of my ability, preserve, protect and defend the Constitution of the United States.” - President Donald J. Trump, January 20, 2017.
"And it's, frankly, disgusting the way the press is able to write whatever they want to write. And people should look into it." - President Donald J. Trump, October 11, 2017.
Fast Eddie B is online now   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 14th January 2018, 04:28 PM   #96
RecoveringYuppy
Philosopher
 
Join Date: Nov 2006
Posts: 7,779
Originally Posted by Fast Eddie B View Post
I wondered that as well.

Two options I see...

1) During the course of the match it gained a new insight into what the best first move was, or...

2) After playing myriad games with itself, it found that even against itself it won more often if it was not 100% predictable in its first move.

Still, a fascinating question.
3) Two or more openings have same likelihood to win.
__________________
REJ (Robert E Jones) posting anonymously under my real name for 30 years.

Make a fire for a man and you keep him warm for a day. Set him on fire and you keep him warm for the rest of his life.
RecoveringYuppy is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 14th January 2018, 08:03 PM   #97
Thor 2
Illuminator
 
Thor 2's Avatar
 
Join Date: May 2016
Location: Brisbane, Aust.
Posts: 3,976
Originally Posted by RecoveringYuppy View Post
3) Two or more openings have same likelihood to win.

If that is the case how does the program make the choice? Can it be subjected to whims?
__________________
Thinking is a faith hazard.
Thor 2 is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 15th January 2018, 01:38 AM   #98
Cheetah
Graduate Poster
 
Cheetah's Avatar
 
Join Date: Feb 2010
Posts: 1,028
I don't think this specific link has been posted.

It has some info concerning the openings AlpaZero discovered during the training and how often it used them.
It does not look like it ever settled on a specific opening exclusively.
It's most popular openings ended up being the English and the Queens Gambit, but it only used them about 17 to 19% of the time at most.*

The future is here – AlphaZero learns chess

There are ten of the games against Stockfish that you can run through at the bottom.

*during training
__________________
"... when you dig my grave, could you make it shallow so that I can feel the rain" - DMB

Last edited by Cheetah; 15th January 2018 at 01:41 AM.
Cheetah is online now   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Old 16th January 2018, 11:49 PM   #99
lpetrich
Muse
 
Join Date: Feb 2007
Posts: 755
Originally Posted by Meadmaker View Post
The thing that would make chess960 a possible exception is that I don't know how much difference there would be in the 960 possible variants. Would it have to treat each one as 960 different games? There is plenty of overlap in play, of course, so I am guessing alphazero would do extremely well. With its apparent dominance of position-oriented chess, it ought to do well, but perhaps it would encounter so many never before seen positions that it would fail, or require a great deal more learning time.
To address that issue, I collected some statistics on Chess960's possible initial positions.

Knights, bishops, and queens are evenly distributed, being as probable in one square as in another. Rooks and kings are an exception, with a complementary distribution.

Rooks: {30, 21, 16, 13, 13, 16, 21, 30}/80
King: {0, 9, 14, 17, 17, 14, 9, 0}/80

Knights, Bishops: 1/4
Queen: 1/8

So for Chess960, AlphaZero could plausibly assume that the king is near the middle of the first rank, and that the rooks are near the edges.


Game complexityWP compares the numbers for several board games: board size, number of states, size of search tree, game length in moves, and branching factor.
  • Tic-tac-toe (3*3): 9, 10^3, 10^5, 9, 4
  • Connect Four (7*6): 42, 10^13, 10^21, 36, 4
  • Checkers (8*8): 32, 10^20, 10^31, 70, 2.8
  • Reversi / Othello (8*8): 64, 10^28, 10^58, 58, 10
  • Gomoku (15*15): 225, 10^105, 10^70, 30, 210
  • Chess (8*8): 64, 10^47, 10^123, 70, 35
  • Connect6 (19*19): 361, 10^172, 10^140, 30, 46000
  • Backgammon (28): 28, 10^20, 10^144, 55, 250
  • Xianggi (9*10): 90, 10^40, 10^150, 95, 38
  • Janggi (9*10): 90, 10^44, 10^160, 100, 40
  • Shogi (9*9): 81, 10^71, 10^226, 115, 92
  • Go (19*19): 361, 10^170, 10^360, 150, 250
  • Arimaa (8*8): 64, 10^43, 10^402, 92, 17281
  • Stratego (10*10): 92, 10^115, 10^535, 381, 21.739
lpetrich is offline   Quote this post in a PM   Nominate this post for this month's language award Copy a direct link to this post Reply With Quote Back to Top
Reply

International Skeptics Forum » General Topics » Science, Mathematics, Medicine, and Technology

Bookmarks

Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Forum Jump


All times are GMT -7. The time now is 10:06 AM.
Powered by vBulletin. Copyright ©2000 - 2018, Jelsoft Enterprises Ltd.

This forum began as part of the James Randi Education Foundation (JREF). However, the forum now exists as
an independent entity with no affiliation with or endorsement by the JREF, including the section in reference to "JREF" topics.

Disclaimer: Messages posted in the Forum are solely the opinion of their authors.