Q&A: Phase 38.1 FederatedAggregator — Configuration & Parameter Tuning

[web3guru888]

Q&A: FederatedAggregator (Phase 38.1)

Configuration Questions

Q1: How to choose between FedAvg, FedProx, FedNova, and Scaffold?

Algorithm	Best For	Key Parameter
FedAvg	IID or mild non-IID data, homogeneous clients	C (client fraction)
FedProx	Heterogeneous systems, variable local epochs	μ (proximal term)
FedNova	Heterogeneous local training steps	τ (normalized weights)
Scaffold	High non-IID, client drift is dominant issue	c (control variates)

Q2: What compression ratio should I use?

• Start with no compression (ratio=1.0) as baseline
• TopK with k=0.01 (1%) works well for large models (>10M parameters)
• Quantization to 8-bit is nearly lossless for most architectures
• For bandwidth-constrained settings, combine TopK + quantization

Q3: How many local epochs (E) before aggregation?

• E=1: More communication, better convergence guarantees
• E=5: Good balance for most settings
• E=20+: Risk of client drift, pair with FedProx (μ≥0.01)
• Rule of thumb: increase μ proportionally with E

Q4: Server-side momentum and learning rate?

• Server momentum (β=0.9) accelerates convergence by ~30%
• Server learning rate η_s=1.0 for FedAvg, tune in [0.1, 1.0] for FedProx
• Use learning rate warmup for first 10 rounds

See issue #772 for full specification and acceptance criteria.