Federated Learning (FL) relies on many distributed clients training a shared global model without sending raw data to the server. However, one of the core challenges in FL is determining which clients should participate in each training round, especially when their local models may be outdated or trained under different conditions. This paper proposes a new way to think about scheduling in FL by introducing the concept of Version Age of Information (VAoI). VAoI extends the traditional idea of “freshness” in distributed systems. Instead of only tracking how much time has passed since a client last contributed, it also captures how far the client’s local model has diverged from the latest global model version. This creates a more meaningful measure of update relevance: a client’s model may be recent in time but still stale in content if it has drifted significantly.
Building on this metric, the authors propose a version age–based scheduling policy in which clients with higher VAoI are selected more often for training. The intuition is simple: by prioritizing clients whose updates are most in need of refreshing, the global model can evolve more smoothly and avoid disruptions caused by extremely outdated contributions. When a client is selected and receives the latest global model, its version age is reset. This semantics-aware perspective highlights a broader shift in FL research—moving from purely time-based notions of staleness to metrics that reflect how meaningful or informative an update truly is. By incorporating model relevance into the scheduling process, FL systems can become more stable, more adaptive, and better aligned with real-world deployment environments where devices vary widely in availability and training behavior.
VERSION AGE-BASED CLIENT SCHEDULING POLICY FOR FEDERATED LEARNING