[ExI] alpha zero

spike spike66 at att.net
Thu Dec 7 22:39:28 UTC 2017


 

 

From: extropy-chat [mailto:extropy-chat-bounces at lists.extropy.org] On Behalf Of Dylan Distasio
Sent: Thursday, December 07, 2017 2:16 PM
To: ExI chat list <extropy-chat at lists.extropy.org>
Subject: Re: [ExI] alpha zero

 

Spike-

>…It didn't actually look at existing records.  It played 44 million new games against itself, and when one opponent one resoundingly, that one was selected as the best overall algo (simplifying a bit).  Basically you have a neural net optimizing for best strategy to win a game based on the rules, and evolving an algo based on continued feedback to the system.  It was sui generis though, no historical games were used.  Dylan

 

 

 

OK this is where my imagination fails.  86400 seconds in a day, to do 44 million games means generating them at a rate of 500 per second.  I don’t see how, using the description given, the software could do that.  Even if it did somehow, I don’t see how it could possibly draw useful inferences from it.

 

There is something I am not understanding.

 

spike

 

 

 

 

 

On Thu, Dec 7, 2017 at 4:50 PM, spike <spike66 at att.net <mailto:spike66 at att.net> > wrote:

 

 

>… On Behalf Of Dylan Distasio
Subject: Re: [ExI] alpha zero

 

Spike-

>…You can read more on what a TPU actually is here if you're interested, but they're basically custom hardware that is good at running neural nets:
https://cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu

>…It took 9 hours to train on 44 million different chess games… Dylan

 

 

Ja OK that might be the missing piece: it looked at 44 million chess games.  I had mistakenly drawn the conclusion it didn’t go to existing records but somehow bootstrapped itself to that skill level.  It didn’t generate the games itself.

OK cool, that might have been explained in the original article in ChessNews but I somehow missed it.

spike

 


_______________________________________________
extropy-chat mailing list
extropy-chat at lists.extropy.org <mailto:extropy-chat at lists.extropy.org> 
http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20171207/8123f21c/attachment.html>


More information about the extropy-chat mailing list