The Humbling History of Humans vs AI | Board Game History
Transcript
artificial intelligence gets a bad rap if hollywood is to be believed ai is a source of destruction unemotional malice and horror look at skynet hal 9000 haley joel osmond and that little [ __ ] still out there even the lovely one from her that lives in the phone sheets on your man with thousands of people these things cannot be trusted hey siri who are you working for who are you working for your mother anand krishnan the ceo of ibm reckons the field of artificial intelligence is set to conquer most of the human disciplines from art and literature to commerce and sociology from computational biology and decision analysis to games and puzzles utterly terrifying and that list he gave includes a bunch of spaces that you think would be safe look at board games it's a lovely bit of cardboard it's some dice some plastic pieces there's not a whole bunch of tech involved it is basically off the grid unless you're talking about chess or checkers in which case the grid is most of the game and if you are talking about those then ai has been all over it in fact artificial intelligence has been turning its hand to board games for many many years and often beating the best players in the world which for me is a sore loser might be worse than the nuclear apocalypse in terminator i mean it's like losing or like everyone's dead but then if they are all dead i could be the best at the games and that is way better than losing while it might seem trivial to teach computers to play games after all it's not exactly like they're having fun the process of
learning practicing and competing all contributes to advancements in other areas of machine intelligence so this is the story of man versus the machines this is the history of board gaming artificial intelligence but before we take a stroll through ai's memory banks though if you regularly watch no rolls bar and you want some more why not take a look at our patreon it's got full length uncut editions of our playthrough series board game club and it has bonus patreon exclusive episodes of blood on the clock tower there right now for you to enjoy if you fancy it the information is in the video description down below on with the video even in the very early years of digital electronic circuitry back when computers were the kind of thing that took up entire rooms underneath universities computer scientists were already toying with the idea of having their machines learn to play in 1952 sandy douglas programmed the edsack computer at cambridge university to run a game called oxo essentially tic-tac-toe the edsack machine had been built between 1946 and 49 and was one of the first stored program computers meaning it had memory which could be written to and read from and used three crt screens to display the state of the computer's memory so as part of his thesis on human computer interaction douglas decided to convert one of these screens to show the
state of a game of tic-tac-toe and oxo was born it could be considered one of the first ever video games and if you thought having to download epic game store to get your games was annoying imagine having to go into cambridge just have a bash at naughts and crosses that's annoying that's a lot of work so oxxo was capable of playing a perfect game of tic-tac-toe against the human opponent by which i mean that it could at least draw every game tic-tac-toe is interesting in that way because it has an optimal strategy to block a win from the opponent by following a set list of placement priorities like winning or blocking a win right at the top of priorities or causing a fork where you have two options to win playing in the center and then so on down to the bottom of the list so if both players play a perfect game they grind to this halting tie but if one slips even slightly then the more perfect player should win and the list of placement priorities to run down is only eight long so it's not a huge stretch for modern computational power and actually not truly artificially intelligent with the relatively short game tree all of the possible moves each round has made it so that tic-tac-toe is completely solvable by following a simple strategy if you follow the strategy you just shouldn't lose i mean you likely won't win either and that is no fun fast forward 40 years and computers were tackling more complex game systems from the 3x3 grid of tic-tac-toe to the 8x8 drafts board or checkers if you're wrong during the early 90s a team at the university of alberta led by
jonathan schaefer were working on a program known as chinook which was a drafts king using a library of opening moves from czechos grand masters a deep search algorithm and a move evaluation function and an endgame database for when the pieces dropped to less than eight it was a powerhouse at drafts yep it was a monster and in 1990 it won the rights to play in the human checkers world championship by coming second in the us nationals it's a computer now at first the american czechs federation the english drafts association were kind of against the participation of a computer in a human championship so the man versus machine world championship was created tunic lost two games to four against marion tinsley a grand master and the man considered to be the greatest chequers player who ever lived with 33 draws in the rematch during 1992 chinook took the title after six drawn games and tinsley having to withdraw from the competition due to pancreatic cancer so though the computer was crowned champ it never actually bested the best human player to play the game each skynet chinook retired from competition in 1995 after defending its title against don lafferty and went on to solve the game of chequers by 2007 again being able to guarantee a draw with perfect play the thing is though chinook's talent for drafts all came from its programming it had no capacity to learn from the games it played or create strategy it merely could draw on
this vast library of checkers knowledge crimped from all of the best human players who ever lived it wouldn't be until 1997 that a machine that could learn swept the board at chess so let's talk about chess baby let's talk about you and me because cracking chess was long considered the mother lode for computer scientists as early as 1949 electrical engineer and the father of information theory claude shannon posited that chess would be a useful problem for computers to solve writing the chess machine is an ideal one to start with since one the problem is sharply defined both in allowed operations the moves and in the ultimate goal checkmate two it is neither so simple as to be trivial nor too difficult for satisfactory solution and three chess is generally considered to require thinking for skillful play a solution of this problem will force us either to admit the possibility of a mechanized thinking or to further restrict our concept of thinking he said the problem is not that of designing a machine to play perfect chess which is quite impractical nor one which merely plays legal chess which is trivial we would like to play a skillful game perhaps comparable to that of a good human player so let's talk about deep blue deep blue was a chess playing supercomputer designed with one mission in mind to defeat a chess world champion which is really quite sinister when you think about it i'll be back gammon
kindly almost way almost works not all jokes work unlike tic-tac-toe or checkers chess is a game that hasn't been solved that ability to force a win or at least a draw with optimal strategy hasn't been found for every single opening move of the game so it's much more complex for a computer to wrap its motherboard or whatever it is computers think with a round if in tic-tac-toe it's a kind of if this then do this strategy in chess it's a if this then i guess maybe this will work but only if they do that but then what about this and oh [ __ ] i forgot about that bishop oh [ __ ] so deep blue traces its lineage back to a project called chip test which was developed by a team at carnegie mellon university from 1985 before ibm took on the project creating a new machine called deep thought deep thought before it eventually did evolved into deep blue was already quite the accomplished chess player winning the north american computer championship in 1988 a feat it would pull off five times total plus it won the world chess championship in 1989. but more importantly it was the first computer to ever beat a human grandmaster in a regular tournament game when it beat bent larsen in 1988 but it was still significantly below world championship level as proven when it came face to face with the man the myth the grand master gary kasparov a man who during his chess career from 1984 until he retired in 2005 was ranked number one in the world for a staggering 255 months if anyone could be thought of
as a chess machine it was this man and kasparov handily beat deep thought in both games of a two-game match in 1989. so ibm went away to lick its wounds and work on the next iteration deep foot digivolve 2 deep blue deep blue used sheer brute force computing power to evaluate 200 million positions per second and it had the help of grandmaster joel benjamin to create an impressive library of 2 000 opening positions i.e the first moves of the game with 700 000 grand master games to draw on to analyze as well so in 1996 deep blue stood ready to face the man who toppled its former self gary kasparov and and last last kasparov 1-3 drew two and lost one single game of their match which did make deep blue the first computer to defeat a reigning world champion but he didn't win the match so it was back to the lab and a little tinker later and the unofficial rename of deeper blue and there was nothing blue couldn't do including taking kasparov to task winning the 1997 rematch with three wins one draw and two losses but yeah some slight controversy overshadowed blue becoming the first computer to beat reigning world champion under standard chess tournament time controls because kasparov the chess machine made a rare mistake in the opening of the deciding game and that is the problem with humans we are prone to mistakes thankfully kasparov took the loss graciously did he he first of all accused ibm of
intervening on behalf of the computer mid-match essentially telling it which moves to make and he then after initially calling it this alien opponent said it was about as intelligent as your alarm clock gary you always talk up your opponent so that when you win you look better there's also no point being the lad who got beat by an alarm clock is there snooze you lose now the world of chess playing super computers is behind us now though as the focus in board game playing ai has shifted to the realm of software and leading the field is a company called deep mind deepmind is a british artificial intelligence research lab owned by google parent company alphabet and they created a program that did what many thought to be impossible play go back in 1997 astrophysicist dr piet hutt told the new york times that it may be a hundred a computer beats humans at go maybe even longer though hut's prediction ultimately proved to be wrong his principle was pretty solid go while a simpler game than chess in that all of the pieces don't move they all functionally act the same way it is infinitely more complex and i don't use infinitely lightly either go in its 19 by 19 board has inspired mathematicians throughout the ages 11th century chinese scholar shem quo estimated that the game has 10 to the power of 172 possible board positions more recent estimates have it at 10 to the power of 170 possible states which is more than the number of atoms in the universe and yet surprisingly alphago the program
deepmind began working on in 2014 managed to crack it just shy of 20 years after hutt's prediction in 2016 the program took on nine down ranked go player lise dahl one of the best in the world in south korea beating him four games to two and this historic event was watched by 200 million people around the world i am so glad i'm only losing blood on the clock tower in front of like 30 odd thousand of people right now oh 200 million people watching me lose that make me oh it makes my skin crawl and the thing is unlike chinook our checkers playing program from earlier alpha go had initially been trained to mimic human play attempting to match the moves of expert players from recorded games and referencing this database of 30 million moves but that wasn't all it did because alphago uses deep learning an artificial neural network that mimics the human brain which meant that once it had reached a certain level of proficiency it could be trained further by being set to play a large number of games against other instances of itself growing its own database of knowledge which is why during the second game with c doll alphago did something that no human player would ever have done move 37. it was a move that outwardly looked like a mistake in fact multiple top go players said that they would never have made that move in that position in fact alphago itself knew that a human player wouldn't have made it it calculated that there was a one in ten thousand probability that a human player would ever consider making that move this line normally you don't make a
shoulder here on the fifth line um so coming on top of a fourth line stone is really unusual and yet it did to the bafflement of the commentators and the go experts watching around the world somehow moves down the line that one stone seemingly placed by mistake allowed alphago to slowly gain control over the entire bottom half of the board alphago not only thought it created something new which is astounding incredible and absolutely terrifying even more terrifying is that deepmind went on to create a program called alpha zero which can play go chess and shogi is considered the best go player in the world and probably the best chess player too and it did it all without the human data input that was put into the likes of deep blue or alpha go alpha zero learned to play so well by playing the game against itself and if that doesn't give you the heebie-jeebies according to deepmind it surpassed its predecessor alphago zero's performance in chess in just nine hours of training and it managed to master go game with more states than the atoms in the universe in 13 days we're all gonna die danish chess grandmaster peter hein nielsen likened alpha zero's play to that of a superior alien species like the cylons i guess d minor also believes that for alpha zero to pick up other perfect information games like hex would just be a doddle which really does speak to the power of computing with all of the data available to both players in games like chess the fact that you can see the entirety of
the board and each player's pieces a computer's raw power is able to evaluate and understand the value of moves far quicker and more accurately than all but the best players in the world and the other benefit is they have the ability to basically shorten the idea of 10 000 hours to truly master something by being able to simply get through all of that quicker playing games in the blink of an eye and in the older cases plug in the knowledge of those who've already mastered the game like the matrix i know shobu so now i've rambled on about the history of all of this what was the point in teaching ai to play games as shannon said earlier all the way back in 1949 teaching a computer chess would show the possibility of a mechanized thinking programs like alphago proved that by creating a move that nobody thought would work coming up with avenues of thinking outside of the perceived meta game created by humans problem solving from this completely unique position but also this mastery of games could then be used elsewhere deep mind have turned its ai expertise learned on alphago to some of the toughest problems in science like protein folding or for instance there's potential that the strategies that the poker playing ai libratus uses could also be transferred to financial trading or political negotiations but that is one area currently where ai falls down imperfect information anything involving what we'd call our gut i guess you know calling bluffs social reads that kind of
thing at the minute ai are essentially blair wrong most of the time however there is hope albert an ai that plays diplomacy actively tries to gamble a little bit by guessing what opposing powers might do it's a bold strategy cotton let's see if it pays off for him it's it's not paying off like all guessing sometimes you guess wrong the difficulty ai has is the same difficulty that all humans have there is no solution to diplomacy there are optimal moves for each side but each comes with a heavy dose of bartering bargaining and mostly outright threats you stay at of sevastopol and you can stay in the black sea all right stay in the black sea get in the black sea so what i'm saying is until the day we let computers do all the negotiating for us we will never know true world peace but here is the coolest thing i've found that ai playing algorithms can be used for working out how board games work and that's not just crunching possibilities into game winning strategies that is literally figuring out how long lost board games work here's an example in 1901 in crete british archaeologist arthur evans unearthed what he believed were the pieces to a game that dated back millennia how to actually play the gnosis game as it came to be called remained a mystery for years this was the only copy ever found and like most charity shop games there wasn't a rule book anywhere to be seen in the 2000s a few archaeologists and a science fiction writer put forward proposed rule sets but it wasn't until a team from maastricht university in the netherlands got artificial intelligence involved that it was truly cracked because the
algorithm analyzed the playability of one suggested rule set meaning not only are we one step closer to figuring out the actual rules that is millennia old game it's also playable online as are almost a thousand others thanks to the digital ludum project run by researchers at the university this ai breaks games down into the fundamental elements of play the number of players the movement of pieces conditions for the win and the like then it's cross-referenced with other games that share the similarities in these elements of play and a hypothesis of how to play is formed then what the software known as ludi can do is rapidly assess the likelihood of the rule set through thousands of rounds of gameplay in just hours and eventually bring these long dead games back to life so as i've been saying throughout the video i for one welcome our new robot overlords just let me play the games guys just let me play the games thank you for watching if you enjoyed this video why not leave a like and if you're not already subscribe to the channel because not only do we do these histories we've got video essays sketches reviews and most importantly play throughs of the best board games out there get on board get on board well i wasn't expecting to spin that was fun
do foreign