Training window size for the neural network

[Massa7ca]

Hello,

I am currently developing my own version of AlphaZero to study the training process, and I would like to clarify some details about the training window size used in the Leela Chess Zero project.

1. What is the typical size of the training window used for the neural network at the current stage (e.g., in terms of the number of games)?
2. Does the window size change depending on the stage of training (e.g., early vs. later stages)?
3. If the window size is fixed, how do you prevent the network from "forgetting" strategies learned in the early stages, especially when the total number of games (e.g., several billion) far exceeds the window size

These aspects are crucial for optimizing the training process in my implementation. I would greatly appreciate your insights!

Thank you for your work and attention!

Best regards,