Distributed Validation Models for Adaptive Wireless ML Updates

The patent segments the global ML model into multiple local ML models, each trained on data from specific radio environments. The network device maintains a model pool containing these segmented local models, and selects appropriate models based on current radio conditions. This segmentation reduces the complexity of any single model while maintaining comprehensive coverage across different environments.

The patent performs preliminary training of multiple local ML models offline during periods when radio conditions are relatively stable. These pre-trained models are stored in a model pool at the network device. When radio conditions change, the system quickly selects and switches to a pre-trained model that matches the current conditions, avoiding the need for real-time training and inference of complex models, thus reducing inference time while maintaining accuracy.

The patent implements local ML model training at user equipment, where each device trains models using its own locally collected radio environment data. This eliminates the need to transmit large amounts of raw measurement data to the network for centralized model training. Only essential model parameters and performance metrics are exchanged with the network, significantly reducing signaling load while maintaining high model accuracy through localized adaptation.

The patent implements periodic model training and updating at user equipment, rather than continuous real-time training. The system collects radio measurements over predetermined time periods, then performs model training or updates at these periodic intervals. This periodic approach maintains model adaptability to changing radio conditions while significantly reducing the instantaneous power consumption and processing burden compared to continuous training operations.