FedYogiServerOptimizer<T>
Models & Types · AiDotNet.FederatedLearning.ServerOptimizers
FedYogi server optimizer — adaptive learning rates with controlled second-moment growth.
Inheritance: FederatedLearningComponentBase<T> → FederatedServerOptimizerBase<T> → FedYogiServerOptimizer<T>
For Beginners
Yogi is similar to Adam but more stable with noisy or sparse updates. While Adam’s second moment can grow unboundedly (causing vanishingly small learning rates), Yogi uses an additive update that only grows the second moment when the gradient magnitude exceeds the current estimate, preventing premature convergence.
How It Works
Reference: Reddi, S., et al. (2021). “Adaptive Federated Optimization.” ICLR 2021.
Constructors
| Constructor | Summary |
|---|---|
FedYogiServerOptimizer(double learningRate, double beta1, double beta2, double epsilon) |
Creates a FedYogi optimizer with the direct-constructor defaults. |
Methods
| Method | Summary |
|---|---|
Vector<T> Step(Vector<T> currentGlobalParameters, Vector<T> aggregatedTargetParameters) |
|
string GetOptimizerName() |