4 ms·

You aren't following AI research, clearly. The top Go player in the world retired saying that he thought he was almost perfect, but then a "superior being" sho

You aren't following AI research, clearly. The top Go player in the world retired saying that he thought he was almost perfect, but then a "superior being" showed him he wasn't even close. I doubt there is a human on the planet that can improve the strength of the best AI at Go. Perfect solutions to games is a 40 year old approach. It's something you learn in school, so you can quickly see that brute-force searching is not feasible on even small problem.

bradknowles 7y ago
MCTS isn't doing full width search. Not by a longshot. What it's doing is using the computer's ability to randomly test various lines of inquiry many, many times faster than a human being can, and then use statistical methods to help rule out the lines that cannot produce anything of value. That's why it's Monte Carlo Tree Search. That said, any sufficiently advanced technology is indistinguishable from magic [0], and with modern Machine Learning techniques we are getting closer to crossing the line over to "magic" in certain fields, such as Chess. [0] Clarke's Third Law, see https://en.wikipedia.org/wiki/Clarke%27s_three_laws https://en.wikipedia.org/wiki/Clarke%27s_three_laws
- panopticon 7y ago
  I don't think they were arguing about search breadth. I think they were arguing that Go AI was already formidable in spite of GP's odd claim that search breadth/depth is the reason why. Chess AI like AlphaZero aren't even close to 100% search breadth/depth but achieve startling superiority.
infinity0 7y ago
It's interesting that AlphaStar was "only" able to get into the top 99.8% of human-level play for StarCraft 2, though. By contrast, AlphaZero was able to surpass human-level play for Go quite easily. One wonders whether the inherent approach in AlphaStar (supervised learning) can only ever hope to imitate the best, but surpassing the previous best seems to still be a human domain for now, at least in SC2.
- ivalm 7y ago
  Go has a much, much smaller phase space than SC2. You can't monte carlo search SC, fundamentally the games are approached differently.
  - infinity0 7y ago
    I know. I was responding to this overall thread, which started: > Unfortunately for the article's point, despite a brief moment of viability, even human+computer chess teams are now far outclassed by computers alone. and it then continued: > The top Go player in the world retired saying that he thought he was almost perfect, but then a "superior being" showed him he wasn't even close. I doubt there is a human on the planet that can improve the strength of the best AI at Go. I was making the point that, even though there have been AI advances since the article was written about chess, defeating its immediate point about chess, these advances have also come with their own caveats. ("Oh this technique that beats chess perfectly can't hope to beat Go". "Oh this technique that beats Go perfectly can't hope to beat StarCraft".) These caveats serve to reinforce this article's broader point - that AI has limitations, that humans can complement, and vice versa.

bradknowles 7y ago

MCTS isn't doing full width search. Not by a longshot. What it's doing is using the computer's ability to randomly test various lines of inquiry many, many times faster than a human being can, and then use statistical methods to help rule out the lines that cannot produce anything of value. That's why it's Monte Carlo Tree Search. That said, any sufficiently advanced technology is indistinguishable from magic [0], and with modern Machine Learning techniques we are getting closer to crossing the line over to "magic" in certain fields, such as Chess. [0] Clarke's Third Law, see https://en.wikipedia.org/wiki/Clarke%27s_three_laws https://en.wikipedia.org/wiki/Clarke%27s_three_laws

panopticon 7y ago
I don't think they were arguing about search breadth. I think they were arguing that Go AI was already formidable in spite of GP's odd claim that search breadth/depth is the reason why. Chess AI like AlphaZero aren't even close to 100% search breadth/depth but achieve startling superiority.

infinity0 7y ago

It's interesting that AlphaStar was "only" able to get into the top 99.8% of human-level play for StarCraft 2, though. By contrast, AlphaZero was able to surpass human-level play for Go quite easily. One wonders whether the inherent approach in AlphaStar (supervised learning) can only ever hope to imitate the best, but surpassing the previous best seems to still be a human domain for now, at least in SC2.

ivalm 7y ago
Go has a much, much smaller phase space than SC2. You can't monte carlo search SC, fundamentally the games are approached differently.
- infinity0 7y ago
  I know. I was responding to this overall thread, which started: > Unfortunately for the article's point, despite a brief moment of viability, even human+computer chess teams are now far outclassed by computers alone. and it then continued: > The top Go player in the world retired saying that he thought he was almost perfect, but then a "superior being" showed him he wasn't even close. I doubt there is a human on the planet that can improve the strength of the best AI at Go. I was making the point that, even though there have been AI advances since the article was written about chess, defeating its immediate point about chess, these advances have also come with their own caveats. ("Oh this technique that beats chess perfectly can't hope to beat Go". "Oh this technique that beats Go perfectly can't hope to beat StarCraft".) These caveats serve to reinforce this article's broader point - that AI has limitations, that humans can complement, and vice versa.