Stories
Slash Boxes
Comments

SoylentNews is people

posted by on Wednesday May 31 2017, @12:32AM   Printer-friendly
from the heavyweight-champion dept.

To say that AlphaGo had a great run in the competitive Go scene would be an understatement: it has just defeated the world's number 1 Go player, Ke Jie, in a three-part match. Now that it has nothing left to prove, the AI is hanging up its boots and leaving the world of competitive Go behind. AlphaGo's developers from Google-owned DeepMind will now focus on creating advanced general algorithms to help scientists find elusive cures for diseases, conjure up a way to dramatically reduce energy consumption and invent new revolutionary materials.

Before they leave Go behind completely, though, they plan to publish one more paper later this year to reveal how they tweaked the AI to prepare it for the matches against Ke Jie. They're also developing a tool that would show how AlphaGo would respond to a particular situation on the Go board with help from the world's number one player.

Source: ArsTechnica


Original Submission

Related Stories

AlphaGo Zero Makes AlphaGo Obsolete 39 comments

Google DeepMind researchers have made their old AlphaGo program obsolete:

The old AlphaGo relied on a computationally intensive Monte Carlo tree search to play through Go scenarios. The nodes and branches created a much larger tree than AlphaGo practically needed to play. A combination of reinforcement learning and human-supervised learning was used to build "value" and "policy" neural networks that used the search tree to execute gameplay strategies. The software learned from 30 million moves played in human-on-human games, and benefited from various bodges and tricks to learn to win. For instance, it was trained from master-level human players, rather than picking it up from scratch.

AlphaGo Zero did start from scratch with no experts guiding it. And it is much more efficient: it only uses a single computer and four of Google's custom TPU1 chips to play matches, compared to AlphaGo's several machines and 48 TPUs. Since Zero didn't rely on human gameplay, and a smaller number of matches, its Monte Carlo tree search is smaller. The self-play algorithm also combined both the value and policy neural networks into one, and was trained on 64 GPUs and 19 CPUs over a few days by playing nearly five million games against itself. In comparison, AlphaGo needed months of training and used 1,920 CPUs and 280 GPUs to beat Lee Sedol.

Though self-play AlphaGo Zero even discovered for itself, without human intervention, classic moves in the theory of Go, such as fuseki opening tactics, and what's called life and death. More details can be found in Nature, or from the paper directly here. Stanford computer science academic Bharath Ramsundar has a summary of the more technical points, here.

Go is an abstract strategy board game for two players, in which the aim is to surround more territory than the opponent.

Previously: Google's New TPUs are Now Much Faster -- will be Made Available to Researchers
Google's AlphaGo Wins Again and Retires From Competition


Original Submission

This discussion has been archived. No new comments can be posted.
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
(1)
  • (Score: 2) by Snotnose on Wednesday May 31 2017, @12:40AM (9 children)

    by Snotnose (1623) on Wednesday May 31 2017, @12:40AM (#518004)

    Can it help diagnose cancer? Look at MRI's and recognize diseases? Look at data and figure out something cosmological? Look at trends and do economics better than those that get lucky?

    I get playing go is a win/no win game, but the other's not only aren't but are more important.

    --
    If you're talking about me behind my back, remember you're in a great position to kiss my ass.
    • (Score: 2) by takyon on Wednesday May 31 2017, @12:50AM

      by takyon (881) Subscriber Badge <reversethis-{gro ... s} {ta} {noykat}> on Wednesday May 31 2017, @12:50AM (#518012) Journal

      It's not untrod territory. See: Watson.

      --
      [SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
    • (Score: 2) by frojack on Wednesday May 31 2017, @12:55AM

      by frojack (1554) Subscriber Badge on Wednesday May 31 2017, @12:55AM (#518015) Journal

      YeaH! Right on.

      Lets put it to a real test. Like cracking everybody's encryption for the NSA, and reading all the email to find out who's got the best weed and where he gets it.

      Seriously, there are very few problems in the real world that this kind of AI can solve that don't involve massive collection and sifting of data.

      Somehow we just can't keep ourselves from inventing Skynet.

      --
      No, you are mistaken. I've always had this sig.
    • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @05:15AM (5 children)

      by Anonymous Coward on Wednesday May 31 2017, @05:15AM (#518116)

      The longterm goal of DeepMind, the team behind AlphaGo, is generalized intelligence. The reason Go was an important test is because it's a game with an impossibly large search space paired with a much slower pace that makes traditional game algorithms like minimax + alpha/beta pruning completely ineffective. But paired with this is the fact that there is also right and wrong answers - and you'll know within a few hours which you chose. Something I think that should be memorialized is how the in the first Go match Lee Sedol, one of the most dominant players in the game, was worried he might lose face if he possibly lost a single game to the machine. He lost 1-5 before the #1 ranked player lost 0-3, 0-6 if we include online games. What AlphaGo did was thought to be something many years to decades away.

      The next step is likely Starcraft 2. This is going to be an incredibly relevant test. The reason is that now we have what is literally an infinite search space paired with an infinite input space. And the AI has to be able to pry into this space in real time. I have 0 doubt they could already trivially develop a branch of the software to outperform humans at MRI analysis. But you've got to think bigger. The idea is, eventually, to be able to informally describe a task to an AI - and have it be able to master that task completely autonomously after being given some examples - similar to how a human learns. Imagine the day when you can simply give it a few million MRIs/results and then have it begin to automatically evaluate any new input MRI you give it - or even formulate conjectures relating issues to MRIs that we may never have considered.

      • (Score: 3, Informative) by takyon on Wednesday May 31 2017, @05:36AM (2 children)

        by takyon (881) Subscriber Badge <reversethis-{gro ... s} {ta} {noykat}> on Wednesday May 31 2017, @05:36AM (#518123) Journal

        Lee Sedol lost 1-4.

        I think Starcraft 2 is effectively infinite, not literally infinite. For example, at the first frame of the beginning of the match, you may have only 3 moves: add 1 or more probes to the Nexus build queue, do nothing, or hit the delete key with the Nexus selected and lose. Once the first probe gets built, you might have 5122 or 262,144 (don't know the real number) places that the probe can be asked to move to, some of them being inaccessible terrain which cause the pathfinding algorithm to find somewhere adjacent to move to instead. The probe can be tasked with gathering minerals, or it can build a building.

        The number of possibilities will grow to larger than the number of quarks in the universe in no time at all, but it's finite as long as there are limits on the amount of units you can have and add to build queues.

        Yes, my post is all nitpick, no insight.

        --
        [SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
        • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @06:42AM

          by Anonymous Coward on Wednesday May 31 2017, @06:42AM (#518150)

          Relevant song. [youtube.com]

          Out analing you there, there's time which does indeed make the game tree infinite, even if the state space is technically (and only technically) finite. Go and Chess are guaranteed to end so long as both players make moves, StarCraft 2 is not. If the server update rate is not specifically guaranteed to be a given value then I think there would also be an argument for infinity since we could get into discrete vs continuous issues.

        • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @04:59PM

          by Anonymous Coward on Wednesday May 31 2017, @04:59PM (#518372)

          To be clear here, you are asserting that Starcraft 2 is literally a more difficult and strategically complicated game than Go is?

          Forgive me if I disagree with you.

      • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @03:46PM (1 child)

        by Anonymous Coward on Wednesday May 31 2017, @03:46PM (#518342)

        I think the next step would be games with incomplete information. In go, as in chess, all information about the game is available to both players; indeed, the position on the board already tells you most of what you need to know (in go, I think it tells you everything; in chess, there's some extra information determining the available moves, like whether the king has been moved before — castling is only allowed if it hasn't —, but again that information is available to both players).

        In the real world, you rarely have the complete information available. Therefore being able to cope well with incomplete information is important.

        • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @06:06PM

          by Anonymous Coward on Wednesday May 31 2017, @06:06PM (#518419)

          A far lesser AI system just completely destroyed [ieee.org] 4 of the best heads up no limit hold'em players in the world. It was a decent sample size and hands were mirrored to minimize the impact of luck. In other words if he ran his kings into aces, he would be simultaneously running his aces into kings against a player playing the mirror setup. At the end, the AI was positive against every player individually. The AI beat the pros with a winrate that's what you'd expect to see in a pro vs strong amateur match.

          I think DeepMind going for Starcraft 2 is going to likely be the greatest AI challenge to date. Incomplete information, real time, infinite search space, and the skill level of the top humans is incredibly high.

    • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @12:35PM

      by Anonymous Coward on Wednesday May 31 2017, @12:35PM (#518231)

      Can it find the answer to life, the universe and everything? :-)

  • (Score: 2) by Kilo110 on Wednesday May 31 2017, @03:20AM (1 child)

    by Kilo110 (2853) on Wednesday May 31 2017, @03:20AM (#518066)

    "Google-owned DeepMind will now focus on creating advanced general algorithms to help scientists find elusive cures for diseases, conjure up a way to dramatically reduce energy consumption and invent new revolutionary materials."

    And most importantly... find even more creepy ways to data mine their users' data.

    • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @02:34PM

      by Anonymous Coward on Wednesday May 31 2017, @02:34PM (#518304)

      Or maybe they'll use it to solve this problem. [xkcd.com] (See the bottom panels)

  • (Score: 0) by Anonymous Coward on Wednesday May 31 2017, @05:45PM

    by Anonymous Coward on Wednesday May 31 2017, @05:45PM (#518397)

    Did it drop the mic on the way out?

(1)