Page 2 of 2

Re: lczero rating

Posted: Mon Apr 02, 2018 9:34 pm
by George Tsavdaris
stavros wrote:
George Tsavdaris wrote:
stavros wrote:correct me if iam wrong but even google Alphazero progress saturated after 700000
steps https://arxiv.org/pdf/1712.01815.pdf#page=4
i cant imagine lczero to match the latests top emgines.
already latest sd dv+cerebelum book is close to aplhazero
What is "steps"?
from : https://arxiv.org/pdf/1712.01815.pdf#page=4

"We trained a separate instance of
AlphaZero
for each game. Training proceeded
for 700,000 steps (mini-batches of size 4,096)
So how these "steps"/"mini-batches" are compared to games?

Re: lczero rating

Posted: Tue Apr 03, 2018 6:16 am
by jkiliani
George Tsavdaris wrote:
jkiliani wrote: It will not be necessary to start from zero once the network stalls. Instead, a larger neural net can simply be trained from existing self-play games, afterward the net can continue to improve.
What is the ratio of time of generating self-play games to training from these games? If it is 10:1 for example then creating a bigger NN and training it again then no harm is done once you have the self-played games.

BUT since these self-played games have been played by a smaller(and weaker) NN, by training from them a bigger NN, doesn't this creates an non optimum procedure?
The ratio of computation power going into self-play to training is much larger than 10:1, more like 50:1 I think.

Bootstrapping a larger neural net from a smaller one has been tested with Leela Zero, and has been very successful there. So there's little cause for concern that this would negatively impact the network in any way.