So last time we were here Artemy finished up the discussion on introducing the beginnings of machine learning and today what we're gonna talk about over the next couple of lectures is a case study in machine learning applied to a specific system so board games have a very long history and AI going all the way back to the 1950s and they were used to kind of drive the initial developments and artificial learning and machine intelligence so why is it why do we care about board games well on the one hand board games are this really isolated system where we know what it means to win we know what all the moves are and we kind of know the consequences of the of those actions unlike in the real world where things would be really complicated and messy and people change and times change and so it's hard to have it a nice replicable system whereas in a board game the rules of chess have been the same for hundreds of years but also board games have been this measure of what people think of as intelligent and whether or not that's true we can see there are many features of a board game that were late to the world world so in particular in this article but from byte magazine in 1978 we can see these features that are really important to playing a game so you need to plan out a strategy to figure what moves you're gonna make and how to execute them you need to plan ahead to figure out what your goals are both short term and long term and then you need to be able to calculate how these moves are going to play out in order to execute a good plan so we can see that while these features might be in their own specific ways in chess there are actually the same kinds of things that we do in real world systems making board games an ideal place to start trying to think about how to make machines do things that seem like they're intelligent computer scientists have been building machines to play board games now for a long time as well so back in nineteen fifty two of the Machine oxo Pam managed to play a perfect game of tic-tac-toe flashing forward in 1995 the computer Chinook became the world champion in checkers now the technically Chinook became the champion in a way that computers can which is that humans are mortal and in fact the best human checkers player player Marian Tinsley actually passed away in the middle of their match and Marian Tinsley was so much better than every other player that all of the sudden Chinook became the world champion but we'll never know if Chinook was better and then many of you may have heard in 1997 deep blue beat garry kasparov - who is the world chess champion at the time so we see that 1952 is tic-tac-toe and then 1995 checkers is beaten and then 1997 chess the computer becomes the world champion so 1998 how were computers doing it go well if you see this board here you might not know how the rules are but the computer got to play 30 moves at the beginning of the game and still lost to a decent human not even the world's best human so you can see that the way the computers are the best player ever in chess but still needed at least 30 30 moves at the beginning of a go game in order to beat top humans so now let's take a minute divergent to talk about the rules of go and we'll talk about how it works so first let's talk about the history so go is very popular worldwide but especially in East Asia and there are over 30 million players throughout the world and in fact there are between four and six tournaments every year that have a prize winnings of over a quarter of a million dollars go is between 2,500 and 4,000 years old and in fact some of the oldest texts that we have about go is the writings of Confucius so go has a very long history as well so in particular it go is a game it's a two-player game so there's two players but it's also deterministic which means there's no randomness so if you play something like poker you don't know what card you're gonna draw when you draw cards whereas in go you play a move and that's that and it's also a game of perfect information which means that both players know everything that's going on so you could also imagine say a game about war where there's a fog of war so maybe you have pieces but I don't know where those pieces are and so gantt go is not that it every everyone knows everything which makes it a very nice system for machine learning because we don't have any hidden information so the way it works is that I said there are two players black and white and they alternate moves the board is 19 points by 19 points and players alternate playing on the intersections the next idea is this idea of capturing if you have a group of stones that are connected by points and they get surrounded entirely by stones of the opposite color those stones are removed from the board now black and white play back and forth taking off stones if necessary if they get captured and at the end of the game we finally see who is surrounded the most territory and so what we do is we count up the territory surrounded by both players and then we count up the number of stones that you have captured and whoever has the most points wins so that's not all of the rules ago but it's almost all the rules to go and we'll hopefully give you enough to moving forward to give you some sense of how the game plays so going back to computer go as I mentioned back in 1998 computers are doing really really poorly and then around 2000 we finally had one of our first breakthroughs which was talk later but in 2004 a computer loss to a professional with a nine stone handicap which is still a big handicap but it's it's much less than 30 stone handicap and then by 2010 we had decreased another two stones so a computer beat of professional with a seven stone handicap and then another five years on five years later in 2015 a computer beat a professional with a five stone handicap so from 2004 until 2015 so eleven year span we decreased four-star handicap couple months later and March of 2016 this program alphago defeated Lisa Dole who is the one of the top professions at the time with a zero stone handicap so it took us 11 years to decrease four stones and then in like three months it went from five stones to beating one of the top professionals and then a couple of months later the second version of the program alphago master beat the top professional coach a three to nothing so the computer didn't even lose and in qje was one of the top players at the time and then recently a couple months ago alphago zero came out and defeated the alphago master version eighty nine to eleven so it took us eleven years to break four stones but now in the last year and a half we've managed to like beat the top players and then beat programs that have beat the top players so we've had this amazing level of progress