I need to run experiments using as a baseline the most recent and best aggregation method considered state-of-the-art implemented in Flower.
Could someone please tell me what it is?
Thank you.
Hi @gestefane
There isn’t a single “best” or universal SOTA aggregation algorithm, as the optimal choice strongly depends on the targeted use case (e.g., level of data heterogeneity, system heterogeneity, robustness requirements, personalization, etc.). Different methods are considered SOTA under different assumptions and evaluation settings.
For a general and widely accepted baseline, FedAvg is still the standard choice and is implemented in Flower. Of course, it’s generally good to include more baselines based on your expected scenarios.