Implemented aggregation is different from the description in McMahan 2016 paper

[vaseline555]

Implemented aggregation is different from the description in McMahan 2016 paper

[bokun-wang]

Thanks for the clarification.

This paper also looks relevant, which suggests specific values for the server-side step size eta.

[vaseline555]

Dear @bokun-wang,
Hi! Thank you for your interest in my repository. 😃

Exactly as you understood, it is a server momentum, which is firstly proposed in FL by Hsu et al., 2019, coined as FedAvgM.
This concept has also been adopted into other FL researches, e.g., Dinh et al., 2022 (pFedMe).
Since vanilla aggregation mechnism is the special case of the momentum verison (i.e., $\eta=1$), I intentionally implemented in that way.
Please correct me or PR, if I made an wrong statement or incorrect implementation.

BTW, I think I should explictly add fedavgm option to the choice of algorithm argument that forces the momentum constant (i.e., $\eta$) to be positive.
Thank you!

Best,
Adam