Artificial Intelligence by Melanie Mitchell

Author:Melanie Mitchell
Language: eng
Format: epub, mobi, azw3
Publisher: Farrar, Straus and Giroux

“Without Human Examples or Guidance”

Unlike supervised learning, reinforcement learning holds the promise of programs that can truly learn on their own, simply by performing actions in their “environment” and observing the outcome. DeepMind’s most important claim about its results, especially on AlphaGo, is that the work has delivered on that promise: “Our results comprehensively demonstrate that a pure reinforcement learning approach is fully feasible, even in the most challenging of domains: it is possible to train to superhuman level, without human examples or guidance, given no knowledge of the domain beyond basic rules.”4

We have the claim. Now let’s look at the caveats. AlphaGo (or more precisely, the AlphaGo Zero version) indeed didn’t use any human examples in its learning, but human “guidance” is another story. A few aspects of human guidance that were critical to its success include the specific architecture of its convolutional neural network, the use of Monte Carlo tree search, and the setting of the many hyperparameters that both of these entail. As the psychologist and AI researcher Gary Marcus has pointed out, none of these crucial aspects of AlphaGo were “learned from the data, by pure reinforcement learning. Rather, [they were] built in innately … by DeepMind’s programmers.”5 DeepMind’s Atari game-playing programs were actually better examples of “learning without human guidance” than AlphaGo, because unlike the latter they were not provided with the rules of their game (for example, that the goal in Breakout is to destroy bricks) or even a concept of the “objects” relevant to the game (for example, paddle or ball) but learned exclusively from the screen pixels.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.