Featured Application
The evaluated application is a simplified synthetic scheduling benchmark. It tests whether federated weight learning improves dispatching under explicit assumptions. The new results do not demonstrate an advantage over centralized or local-only learning. Physical safety, industrial calibration, privacy protection, and equipment cost savings are outside the demonstrated scope.
Abstract
We rebuild the evaluation of a federated LSTM weight predictor around a corrected dimensionless scheduling cost and an explicit teaching procedure. This Python 3.12 reconstruction replaces the previous unverified MATLAB figures; it is not a reproduction of the unavailable original implementation. Four virtual workshops generate heterogeneous data, and a finite search over 35 candidate weight vectors supplies causal, myopic targets. Across 30 independently regenerated training and testing seeds, ten methods are evaluated under normal and burst arrivals, giving 2400 workshop-level runs. Mean deadline-or-horizon failure rates for the federated predictor are 4.54% and 18.05%, respectively. Paired comparisons with centralized, local-only, fixed-weight, and memory-free alternatives do not show a significant advantage after multiplicity adjustment. Under burst arrivals, nearest-distance dispatch achieves 15.54% failure and outperforms the federated predictor. The study therefore provides a reproducible negative finding for this configuration, rather than evidence of federated superiority. Labels, checkpoints, task streams, event logs, source code, and uncertainty estimates accompany the reconstruction. Simplified traffic and energy models, a finite teacher, limited workloads, and the absence of physical or privacy validation restrict generalization.