AsyncFedEDTrainer<T>

Models & Types · AiDotNet.FederatedLearning.Trainers

Implements AsyncFedED — Asynchronous FL with Entropy-Driven client scheduling.

Inheritance: FederatedLearningComponentBase<T> → AsyncFedEDTrainer<T>

For Beginners

In async FL, the server doesn’t wait for all clients — it processes updates as they arrive. But which clients should train next? AsyncFedED uses an information- theoretic approach: it estimates each client’s “information gain” (how much the global model would improve from that client’s update) using entropy of the client’s local loss distribution. Clients with higher entropy (more uncertain predictions, more to learn from) are scheduled first. This prioritizes the most informative clients, converging faster than random scheduling.

How It Works

Scheduling:

Reference: AsyncFedED: Entropy-Driven Scheduling for Asynchronous Federated Learning (2024).

Constructors

Constructor Summary
AsyncFedEDTrainer(double stalenessDecay, double explorationBonus, int selectionBudget) Creates a new AsyncFedED trainer.

Properties

Property Summary
IReadOnlyDictionary<int, double> ClientEntropies { get; } Gets a snapshot of the current entropy estimates for all tracked clients.
double ExplorationBonus { get; } Gets the exploration bonus.
double StalenessDecay { get; } Gets the staleness decay factor.
int SelectionBudget { get; } Gets the selection budget.

Methods

Method Summary
Dictionary<string, T[]> AggregateWithEntropyWeights(Dictionary<int, Dictionary<string, T[]>> clientUpdates, int currentRound) Aggregates client updates with entropy-weighted contributions.
List<int> SelectClients(IReadOnlyCollection<int> availableClients, int currentRound) Selects the most informative clients for the next round.
void UpdateClientEntropy(int clientId, double[] classLosses, int currentRound) Updates the entropy estimate for a client based on their local loss distribution.