Peer-to-Peer Feature Exchange for Edge Inference

A centralized model training service acts as an intermediary between network elements and forecasting models. This service collects telemetry data from multiple network entities, trains specialized forecasting models for different network configurations, and distributes them to edge devices. This mediator handles the complexity of network-specific feature construction, allowing edge devices to execute forecasts without dealing with the underlying complexity.

The forecasting system is segmented into distinct components: a centralized training service that handles model development and a distributed edge inference system that executes forecasts. This segmentation allows the complex model training process to be separated from the simpler inference process, reducing the complexity burden on edge devices while maintaining predictive capabilities.

The system performs preliminary actions by pre-training forecasting models with comprehensive telemetry data from multiple network entities during off-peak periods. These pre-trained models are then deployed to edge devices, where they can make rapid forecasts without needing to collect and process data from multiple peers in real-time. The heavy data collection and processing work is done in advance.

Instead of having each edge device collect and process telemetry data from multiple peers, the system creates copies of pre-trained forecasting models and distributes them to various edge devices. Each device executes local copies of these models using its own telemetry data, eliminating the need for real-time inter-device data collection while maintaining forecasting accuracy.