There are not many sources online, but one reference from January says of Leela Zero (LZ) that:

The strength depends on the hardware and on thinking time, but from the thread "LeelaZero adventures on Fox", and from petgo3's rank on KGS, I guess that on medium hardware, and with relatively fast time settings, LZ is about professional strength but not top pro

At the time this was published LZ's best network had had ~11.75 million training games.

In contrast, AlphaGo Zero (AGZ) had a total of ~29 million training games. Assuming that the number of training games were spaced out evenly, the number of LZ's training games by January would correspond to about day 14 of AGZ training.

But by day 14 of training AGZ was very close to AlphaGo Master's level, which is way above top pro level.

This doesn't add up, LZ was made mimicking AGZ so the growth rate should be similar, but this makes it seem like it is considerably slower.

What factors could be causing this? Is it the hardware (like the Tensor units?) used for the runs or the depth with which the games are analyzed? Is it something having to do with the implementation of the neural network itself? Or... maybe... is it just that LZ has very bad luck (Averaged over millions of games this one seems very unlikely, but it's a possibility)?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.