5 ms·
> Go does not have openings. There are 381 opening moves in Go, but really only 96 because of symmetry. 96 (opening moves) x 380 responses x 379 x 378 x 377 ==
by dragontamer 7y ago
> Go does not have openings.
There are 381 opening moves in Go, but really only 96 because of symmetry. 96 (opening moves) x 380 responses x 379 x 378 x 377 == ~2 Trillion positions after 5 ply.
These 2-trillion positions will easily fit in a 16TB hard drive for $400. That's 8-bytes per position, so you probably can get there with more symmetries and some compression applied.
----------
You're thinking too much like a human. There's no Go-openings in the age of Human-Go. But in the age of Cyborg-Go where 16TB hard drives are allowed, we can begin to exhaustively build openings.
We even have a super-human AI that can automatically, and algorithmically, explore this opening book. We can build AWS-instances with V100 Tensor cores to use neural-nets to explore all of these positions now.
-------------
> If you have alpha-go play itself a thousand times, it is unlikely that by move 10 you will wind up with the same board position twice.
Alpha-Go doesn't seem to implement much randomness at all into the moves it plays. The source of randomness is in time-controls (AlphaGo may choose MoveX before 30 seconds of analysis, or MoveY after 30 seconds of analysis), but this is a fairly constrained number of moves.
Play alphaGo by itself a thousand times, at precisely the same time MCTS-controls (say: 1-million nodes visited in the MCTS tree), and it will probably play the same game 1000 times in a row.
This makes AlphaGo extremely prone to opening database "attacks". Which is why I am using opening books as an easy example for how to beat a particular AlphaGo network. At least, until AlphaGo updates its algorithm for more random play.
If the goal is "Beat AlphaGo" in a game, then the opening book construction is far, far simpler. Even with random elements (ex: AlphaGo picks randomly from the top 10 best positions it generates), that is far more constrained than a full 381 x 380 x 379 x ... style opening book.
- btilly 7y agoYou're thinking too much like a human. No, you're thinking too much like someone who understands chess but not Go. Suppose that we built an opening book with a trillion reasonable courses of action on it. Each one analyzed well. As you have discovered, you will only go a few ply into the game. And all of the positions that you will be directed towards will have only a small edge. Instead put a tiny fraction of the computing power necessary to build this book into self-training. You will get a better internal model and therefore a significantly stronger computer player. (That is how alpha-go was built in the first place.) This option will produce much better results for far less effort, and again leaves no role for a human to do anything useful. The fact that a memorized opening book is useless has nothing to do with human vs computer vs cyborg. It has to do with the characteristics of the game. In chess, it is useful to memorize openings and both computers and humans do it. In Go it is a waste of energy, and it is a waste for both computers and humans.
- dragontamer 7y ago> Instead put a tiny fraction of the computing power necessary to build this book into self-training. Do you believe that AlphaZero could continue to improve dramatically with another 6-months of training? Or if it can improve at another 6-months after that? At some point, the network will reach a local maximum, and it will be unable to improve beyond that. Characterizing AlphaZero's moves through big-data analysis is innately going to become useful as self-training plateaus. Even Google wasn't able to get more than a few months of training in before the plateau. At which point, it will be more reasonable to characterize the weaknesses of the network and build an opening book. Avoid the positions that the network was unable to learn about. Etc. etc. Opening books, at a minimum, would grossly improve AlphaZero's play at competitive levels. Anyone with an opening book of AlphaZero's mistakes will be able to push AlphaZero into a mistaken position.
- jmcgough 7y ago> Anyone with an opening book of AlphaZero's mistakes will be able to push AlphaZero into a mistaken position. This only works if AlphaZero never retrains on previous games.
- btilly 7y agoAlphaGo spent months training and continued to improve for the whole time. Its improvement slowed, just as it takes more work for humans to get from master to international master than it does from D-level player to C-level player. But it did not stop improving. And then AlphaZero was better than AlphaGo after around a day of self-training. Furthermore you are arguing for an opening book without considering how small an advantage an opening book would be. As I have said repeatedly, an opening book takes a tremendous amount of work to generate, will only go a few ply in, and the positions it directs you to will only be a tiny bit better. Therefore for the foreseeable future, more training and better algorithms will produce better results than trying to create an opening book. Theoretically this could change. But that day will not be today or this year. I would be astonished if it happened during this decade. I would be surprised if it happened in a lifetime. Your proposed approach is an excellent one for many games. But not for Go.
- mlyle 7y agoSo you use perfect analysis to know, 5 moves deep, how to get the biggest advantage possible against AlphaGo. This is already pretty hard to analyze so deeply. Then you wring out a tiny advantage-- a fraction of a stone. It's a small benefit compared to building other parts of understanding of the game.
- EsssM7QVMehFPAs 7y ago> Alpha-Go doesn't seem to implement much randomness at all into the moves it plays. The source of randomness is in time-controls (AlphaGo may choose MoveX before 30 seconds of analysis, or MoveY after 30 seconds of analysis), but this is a fairly constrained number of moves. The basis of reinforcement learning algorithms is the exploratory nature of learning due to the initial application of largely random moves. Only after some time the agent is given confidence into his learned ability and grafually moved into a more deterministic behaviour mode. This is the exact opposite of your statement. Star Craft players have noted that the fleet of different AlphaStar instances training in ensemble observed very different behaviour due to this property of RL.
- kerkeslager 7y agoGosh, it's sure clear that you don't know what you're talking about. At 5 ply, the complexity of the game hasn't started in any meaningful way. In a typical game, that's 4 corner moves, and then one of a: an approach to a corner (kakari), b: an enclosure (shimari), c: a wedge (waruichi) or d: creating a side framework (such as the Chinese fuseki or Sanrensei). There are some odd opening such as tengen, or corner-corner-corner-kakari which typically turns into a sente fight, but 99% of games will fall into the aforementioned pattern. The database you describe is about as useful as a database of amateur games, since most games, including AlphaGo's games, follow just a few basic openings that early, and even amateurs can play these first few moves "correctly". Even if you get out to 10 ply you're still only getting partway into a single joseki sequence, often leaving three whole corners of the board which haven't even been approached, so this database still isn't very useful. Incidentally, your numbers are also wrong. Symmetry reduces the first move to 55 possibilities, not 96, and there are 361 points on the board, not 381.