10

I was reading a UK Government paper on AI and technology and found this section to be slightly unbelievable.

when Google DeepMind’s AlphaGo played Lee Sedol in March 2016 (see paragraph 3), the machine was able to beat its human opponent in one match by playing a highly unusual move that prompted match commentators to assume that AlphaGo had malfunctioned.81 AlphaGo cannot express why it made this move and, at present, humans cannot fully understand or unpick its rationale.

Is it true that people can't work out the rationale behind the move in question?

Brythan
  • 10,162
  • 5
  • 46
  • 53
Luke
  • 209
  • 1
  • 3
  • [Welcome to Skeptics!](http://meta.skeptics.stackexchange.com/questions/1505/welcome-to-new-users) What sort of evidence would you like to see in an answer? I am not sure how I would address this one. – Oddthinking Mar 01 '17 at 12:20
  • 10
    "For Move 37, the probability was one in ten thousand. In other words, AlphaGo knew this was not a move that a professional Go player would make. But, drawing on all its other training with millions of moves generated by games with itself, it came to view Move 37 in a different way. It came to realize that, although no professional would play it, the move would likely prove quite successful."-https://www.wired.com/2016/03/googles-ai-viewed-move-no-human-understand/ – pericles316 Mar 01 '17 at 13:18
  • 6
    I'm voting to close this question as off-topic because it properly belongs on a programming-related site – Avery Mar 01 '17 at 14:20
  • This question certainly doesn't belong to stack overflow. It might be possible to tweak the up to progammers' standards, not sure. – John Dvorak Mar 01 '17 at 18:19
  • 6
    I think there's a bit of uncertainty here about what it means to "understand" a program trained with a deep neural network. The actual logic that leads to decisions in a deep learning setting is completely opaque: if anyone understands AlphaGo's rationale at all it's just because they're assuming that it's mimicking the behavior of an expert player. In the case of this example, it's just that the program behaved in a unique way and made a move a human expert probably wouldn't have. – ThatGuy Mar 01 '17 at 19:57
  • 5
    That said, AlphaGo did use what's called a game tree, which basically means it had the ability to simulate chains of actions to see what would happen. If the game trees generated during the game were saved somewhere, then it might hypothetically be possible to look through them and see what moves/outcomes Alphago was considering just before it acted. But I doubt anyone did that to any serious degree. The fact is, most of Alphago's supremacy came from it's neural network, and there's no way to know - at a fundamental level - why that preferred some moves to others. – ThatGuy Mar 01 '17 at 20:05
  • Seems like that link @pericles316 confirms humans cannot fully unpick it's rationale due in part to the scale of information that has been processed in reaching that particular decision. – Luke Mar 02 '17 at 10:02
  • 1
    @Luke-Its not completely true since there was a move 78 in another game which was not predicted by Alpha go. "Lee won his only game of the tournament with a flash of brilliance, one that AlphaGo had dismissed as a realistic human possibility In an unexpected irony, it was the same probability assigned to its own dazzling move 37 in the second game. And like Lee then, it couldn’t come back from the shock now." – pericles316 Mar 03 '17 at 14:01
  • @Avery - I agree. This is a great question, but belongs on CompSci or SO or maybe Math – warren Mar 03 '17 at 18:14
  • There's a beta site for AI -- this question would fit great in there. – Alpha Mar 07 '17 at 12:35
  • Happy enough for this to be migrated if it's a better fit elsewhere. :) – Luke Mar 07 '17 at 12:41
  • @pericles316 Although in fairness, AlphaGo was severely detuned for that game. – Charles Mar 16 '17 at 06:24

0 Answers0