66512 net

[pawel....@gmail.com]

This net is  first one with Elo increased from around 3000 to 3050. It is also much bigger in size than previous ones, 160Mb vs 90Mb. Can smbdy explain please the reason for such a significant jump in size and elo at the same time ? 

[glbchess64]

66512 is a 384x30b net, 66511 a 320x24b net. T60 received a new initialisation with J92-320 which is stronger than 66511 at fixed nodes (about 50 elo at fixed low nodes).

[pawel....@gmail.com]

Wow. So ...I will run some tests 66512 or later vs 66511 to see what the elo difference 1min +2 sec. I will post the result here tomorrow I think.

[hypermoder...@gmail.com]

Can you just create an even bigger net just using all the training games when the main net progress plateau ? and repeat the process creating a bigger and bigger NN without having to restart from scratch the training ?

[hypermoder...@gmail.com]

I think of this like a growing brain ... does what I say work ? is it what your doing ?

[pawel....@gmail.com]

You have to take into consideration that growing the net slows down nps significantly. So as a matter of fact it can slows down the speed to the degree that it will be useless without specialized computers. On the other hand, it improve playing strength with only 1 ply depth. So for example bigger net is better then corresponding smaller one when no calculations are involved at all, playing just by pure intuition. 

[DBg]

Yes, all the learning would be about the position evaluation function.  If the number of layers were on par with the "ultimate" evaluation function, there would be no need for node computations, all would be sorted out from a very precise evaluation function, with the ultimate intuition level  (à la 42, or maybe just à la intrinsic topological complexity of full chess position set or its embedding in the space spanned by the appropriately big neural net basis of functions).  That's a hypothesis for rephrasing the above.... no tomatoes please, i am just doing a hit and run, and reading the above, my itch got hard to repress.