Predictive Transmission Path Selection for Low-Latency Packet Routing

The patent applies preliminary action by using machine learning models to predict future transmission path parameters before actual data transmission occurs. The system trains models on historical network data and uses these trained models to forecast RTT, bandwidth, and other critical parameters for upcoming transmissions, enabling proactive path selection rather than reactive selection based on delayed measurements.

The patent introduces machine learning models as intermediaries between raw network measurements and transmission path selection decisions. These models process historical measurement data and generate predicted future performance metrics, serving as a mediator that transforms past observations into actionable predictions for future path selection, thereby improving accuracy while maintaining real-time decision-making capability.

The patent applies self-service by implementing automated machine learning model training and evaluation systems that continuously monitor network conditions, train models on collected data, evaluate model performance, and deploy updated models without manual intervention. This automation reduces the operational complexity of managing multiple transmission paths while maintaining high reliability through continuous adaptation to changing network conditions.

The patent utilizes parameter changes by dynamically adjusting transmission path selection based on predicted parameters such as RTT, bandwidth, and packet loss probability. The system monitors multiple parameters simultaneously and uses machine learning models to forecast their future values, enabling flexible adaptation to changing network conditions while managing complexity through focused parameter monitoring rather than comprehensive system oversight.

The patent applies partial action by implementing incremental and online learning approaches where machine learning models are trained progressively on streaming network data rather than requiring complete retraining on entire historical datasets. This allows the system to maintain prediction accuracy while significantly reducing training time and computational resource requirements by updating models with recent data samples in a continuous learning process.

The patent implements periodic action by scheduling model training and evaluation at regular intervals based on accumulated data quality and quantity thresholds. Rather than continuously training models on all available data, the system periodically retrains models when sufficient new data has been collected, balancing prediction accuracy improvement with computational efficiency and reducing overall training time while maintaining model performance.