Base Station UE Selection for Federated Learning Scheduling

The patent segments the UE selection process into multiple stages: initial scheduling indicator calculation, threshold-based filtering, and iterative refinement. By dividing the large set of all UEs into smaller subsets through sequential filtering steps, the system maintains model quality while reducing the number of UEs participating in each training round, thus decreasing communication overhead and convergence time.

The patent dynamically adjusts the channel gain threshold parameter based on system conditions, historical performance, and convergence requirements. By changing this threshold parameter adaptively, the system optimizes the balance between selecting enough diverse UEs for model quality and limiting the number to reduce communication overhead, directly addressing the contradiction between reliability and time loss.

The patent applies local quality by evaluating each UE's individual channel condition and assigning different selection probabilities or thresholds based on their specific characteristics. Instead of uniform treatment, UEs with better channel conditions receive higher priority, while those with poor conditions are selectively included only when they provide unique data value, thus maintaining data diversity without compromising communication reliability.

The patent implements partial action by selecting only a subset of UEs that meet the channel threshold criteria for each training round. Rather than including all potentially diverse UEs, the system applies a filtering mechanism that includes just enough UEs with adequate channel conditions to achieve sufficient data diversity, thereby maintaining communication reliability while preserving adaptability.

The patent performs preliminary calculations of scheduling indicators during idle periods or based on pre-measured channel state information before the actual UE selection decision is required. By pre-computing these indicators and storing them, the system reduces the real-time computational burden at the base station while maintaining accurate selection based on pre-analyzed data, thus balancing precision with power consumption.

The patent enables UEs to self-report their scheduling indicators or pre-process their own selection criteria based on local measurements. This shifts part of the computational burden from the base station to the UEs themselves, allowing the BS to make selection decisions with less computational overhead while maintaining accuracy through the UEs' own assessments of their channel conditions and capabilities.