Layman's question: what is the difference between Leela's networks?

[Pedro]

Hello guys

As the title of the topic says, I would like to know what are the differences between Leela's normal networks, the ones we found here on this link: https://training.lczero.org/networks/1, and the Jhorthos' networks, Sergio Vieri etc?

[dkappe]

Hello guys

As the title of the topic says, I would like to know what are the differences between Leela's normal networks, the ones we found here on this link: https://training.lczero.org/networks/1, and the Jhorthos' networks, Sergio Vieri etc?

There are essentially 4 differences:

1. Architecture
2. Size
3. Data
4. Training

1. Examples of architecture might be the addition of the MLH (moves left head) or the WDL vs classical value head.

2. Size would generally be width and depth, so a 128x10 (128 filters and 10 blocks) “small” net vs a 256x20 “medium” net.

3. A network, as it is being trained, generates self-play games. This is the data that is used to train networks.

4. Basically supervised vs reinforcement. Reinforcement learning with leela means the current network generates a little data, trains a little bit to produce a new network, generates a little more data, gets rid of the oldest data, trains a new network, rinse, repeat. The jhorthos and before him Sergio networks were trained on a big batch of the reinforcement data, but often at different sizes. If you squint, this is the “supervised” training.

There’s lots of details and history I’ve left out, but that’s the general idea.