I'm more interested in the fact that they clearly consider the AI to be on par with Lee Sedol, and halfway through they started referring to AlphaGo as 'he' rather than 'it'.
Everything I've read about the Deep Blue games described the computer playing in a very mechanical way, quite differently to what you'd expect from a human grandmaster. AlphaGo seems to play qualitatively very much like a human player.
Unfortunately, while I could follow the tactical aspect, I'm not familiar enough with the game to understand the grand strategy, so it was a bit of an odd experience-- something complicated was happening, but even being familiar with neural networks and reinforcement learning, I couldn't really appreciate what was going on.
In contrast, Deep Blue was an impressive machine, but I understand that a combination of game tree search with position evaluation could allow you to evaluate each move in enough detail to decide whether it was "good" or "bad".
Here, we have a massive neural network developed by one of the companies on the cutting edge of both research and hardware to produce something that is able to beat even the best humans in a game with a much higher branching factor.
I can imagine doing a primitive sort of tree-search by hand with chess using Deep Blue's evaluation heuristics; doing even a few forward propagations through AlphaGo's value network using pen-and-paper would probably consume more time than was allotted for the match.
4 comments
[ 4.3 ms ] story [ 16.2 ms ] threadhttps://news.ycombinator.com/item?id=11250806
Everything I've read about the Deep Blue games described the computer playing in a very mechanical way, quite differently to what you'd expect from a human grandmaster. AlphaGo seems to play qualitatively very much like a human player.
Unfortunately, while I could follow the tactical aspect, I'm not familiar enough with the game to understand the grand strategy, so it was a bit of an odd experience-- something complicated was happening, but even being familiar with neural networks and reinforcement learning, I couldn't really appreciate what was going on.
In contrast, Deep Blue was an impressive machine, but I understand that a combination of game tree search with position evaluation could allow you to evaluate each move in enough detail to decide whether it was "good" or "bad". Here, we have a massive neural network developed by one of the companies on the cutting edge of both research and hardware to produce something that is able to beat even the best humans in a game with a much higher branching factor. I can imagine doing a primitive sort of tree-search by hand with chess using Deep Blue's evaluation heuristics; doing even a few forward propagations through AlphaGo's value network using pen-and-paper would probably consume more time than was allotted for the match.