New AlphaGo Zero reinventing Go without data

Hello,
Have we learnt anything technically about AlphaGo Zero implementation ?
(new version of AlphaGo, not using human data sets, overbeating previous version, while reinventing Go…)
Lots of press but mainly copies of official communication.

It utilizes a form of reinforcement learning which is no covered in these courses.

Yes, thanks, that’s why I am asking, as applied (and very efficient) techniques are apparently quite new and state of art :stuck_out_tongue:
I am starting fast.ai courses soon, but I wanted to have this in mind too ^^
But maybe it has to be firstly trivialized/worked/conceptualized before being accessible to mere mortals o_o

The techniques you will learn in this course are equally on the cutting edge, it is just they are from different domain :slight_smile:

Sort of like you would be comparing being a world class piano player vs being a world class chess player. Both equally good but different domains :slight_smile:

If you want to built more of a conceptual understanding on what the difference is you might want to google supervised learning (what we do in this course - image recognition, sentiment analysis of text) vs reinforcement learning (computer learning to play atari games).

Well, either way - enjoy the course, I think you will have a blast! :slight_smile: Doing the course myself at the moment and having a great time :slight_smile:

1 Like

Here is a very nice single image overview of the architecture of AlphaGo Zero

1 Like