Wireless Congestion Detection via Machine Learning Inversion

Instead of assuming all packet losses are due to bad channel conditions and reducing rates accordingly, the patent inverts the approach by using machine learning to distinguish between collision-induced losses and channel-induced losses. The system learns from historical data to identify the actual cause of packet loss and only reduces transmission rates when channel conditions are truly poor, thereby avoiding unnecessary throughput degradation from false positives.

The patent changes the parameter used for rate adaptation from a simple packet loss counter to a machine learning-based classification output. The system transforms multiple input parameters (packet loss patterns, channel quality indicators, collision metrics) into a refined decision parameter that accurately reflects the cause of packet loss, enabling more precise control of transmission rates.

The patent segments the machine learning system into distinct functional components: a training phase that occurs offline to build the classification model, and an inference phase that runs online with minimal computational overhead. The model itself is segmented into specific features extracted from network traffic (packet loss patterns, inter-arrival times, channel quality) that are fed into a simplified classifier, reducing the complexity of the online implementation.

The patent performs preliminary action by training the machine learning model offline during a training phase using historical network data. This pre-computed model is then deployed for real-time congestion detection, shifting the computational complexity from the online operation to an offline setup phase. The trained model captures complex patterns in advance, enabling simple and fast inference during actual network operation.