How Does DeepMind's AlphaGo Zero Work?

<p>Jan 08, 2018 09:39 PM</p>

<p>

transfer learning.... &nbsp;&nbsp;</p>

<p>

go to minute 5... he explains: &nbsp;policy network vs value network...</p>

<p>

Policy: &nbsp;reduce the # of possible moves</p>

<p>

Value: &nbsp;assign value &nbsp;(eval) &nbsp;the remaining board positions and likelihood of winning from each.</p>

<p>

both were 12 layer CNNs</p>

<p>

used RL &nbsp;in pre-training on games played by humans</p>

<p>

But A-go zero is tabula rasa (no human games at all) and uses combo of policy and value CNN

to play against itself.</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>