A recent BBC article entitled "Google's 'superhuman' DeepMind AI claims chess crown" , discuses DeepMind's recent achievements. While the title is somewhat over the top, it contains interesting information relating to what DeepMind has been up to. It's program, AlphaZero recently won in a chess competition against another AI, Stockfish. It won 75 games as white, 3 games as black, and the other 72 games were draws. This occurred after four hours of training against itself. AlphaZero has also beaten other AI in the Japanese game of Shogi, and a previous version of itself at go after similar hours of training. It seems DeepMind is after all the board game titles. As another article points out, this AlphaZero is more versatile than other programs, as it is not specialized to a specific task. However, this versatility came at a cost, as it used a massive amount of processing power, 5000 custom processors. This article really emphasizes how AlphaZero was given very little help while learning the games, merely the rules of the game, and then it trained for a ridiculously small number of hours against itself. Google hopes this will help DeepMind get closer to a general AI, to help with more complicated problems.
What most interests me about this story is that it is even a competition. Granted, almost half of the chess games were draws, but that still seems to low to me. If both AI are processing all the possible moves then both AI should never lose, and thus they would always tie. This is most true in chess, which AI has been better than humans for a long time. I would not think that having more processing power or a better algorithm would be that much of an advantage. I suppose the AI in my mind were not as strong as the AI in reality, but I think that eventually, even soon, there will be no more room to improve. The programs will be fast enough that they can explore the whole tree and expand all the states, and thus an "Ideal" game will be played, with neither making a mistake. I suppose there could even be multiple of these games. I am curious to see what these games would be, but the sport will be a little less exciting.
The other hugely interesting thing fact about this story is how AlphaZero taught itself quickly, and was not told to use cheap tricks by humans. I think this relates well to our debate which occured yesterday, where we discussed how the nature of how something is created may affect whether or not its creator is creative. I know deep mind was mentioned, specifically with some of the images that were referenced by the "yes" team, which was one of their more compelling points. Perhaps a situation like this, where the AI taught itself its strategies, is an even more compelling argument for the creativity of AI, or perhaps there is something else missing. It is difficult for me to land on one side or the other decisively, but stuff like AlphaZero seriously pushes me to the yes side.
image citation
I think there is a foundation for creativity that can be find through this interesting event. However, there seems to a cap of what creativity can be performed. It is a shallow level of creativity if you ask me since it makes creative product in the possibility of given rules. However, in order to achieve high level of creativity then it would need to be general intelligent which I hope we will achieve soon at some point in the future.
ReplyDeleteComputers are *not* able to try *all* possibilities of a chess game. An average chess position might allow about 35 different possible moves. With a branching factor of 35, the search tree grows exponentially at roughly 35 raised to the Nth power. An average chess game might run for about 80 player-moves. Do you have any conception of how big 35 raised to the 80th power is? Maybe this will help:
ReplyDelete3353077164822748841188787625923370388228499573110866243697217455549154293219967491214428267909397618495859205722808837890625
Thus, I don't think we'll see the game be "solved" so that computers can play a perfect game of chess anytime soon!
P.S. I realize it's hard to understand just how big these exponential numbers are. To add more perspective -- imagine if every grain of sand on earth was transformed into a 16-core 3 Ghz computer, and they all simultaneously worked on exploring different chess positions, then it would still take all of these computers 4.7x10^87 YEARS to search all the possibilities, which is about 3.4x10^77 times the estimated age of the universe! Even if all those computer speeds kept doubling every year for the next 200 years, then it would STILL take about a quintillion universe-ages to complete.
DeleteSo you don't think chess for computers will become like tic-tac-toe for humans any time soon? What if processing power continues its current trajectory? Surely, in 20 or 30 years processing power will be such that terabytes of data won't even be a big deal anymore.
DeleteThe idea the DeepMind being a versatile A.I really interesting, and make A.I's creativity seem nearer than ever. Of course it's still need instruction to start learning any of these game, therefore I still don't think it's have creativity just yet.
ReplyDeleteAlso, about chess "ideal" game, I think since white side have a clear advantage, the "ideal" out come would be more like whoever start as white win instead of all draw. But yet, since DeepMind did not loose when it's on black side, it's the clear winner in this case
You might think that the first player in Checkers would have an advantage, but after they eventually "solved" checkers with massive computing power, it turned out that it is a draw.
ReplyDelete