Develop a dynamic model aggregation strategy for Federated Learning
Instruction: Create a strategy for dynamic model aggregation that adapts to changing data distributions in Federated Learning.
Updated
Official answer available
Read the opening below, then unlock the full answer and practical guidance.
I would first define the target population and a fixed-weight baseline. For a proposed asynchronous system, one simple candidate uses an approved positive contribution weight c_i and model-version age s_i: b_i = c_i/(1+s_i), then a_i = b_i/sum(b_j). Combine accepted deltas as w_new = w_current + eta × sum(a_i × delta_i), with a bounded server step eta. Reject incompatible versions or updates beyond a declared staleness limit; if no acceptable update remains, keep the current model.
At scheduled checkpoints, I...