Estimating TCP throughput: A neural network approach

We address a neural network approach for modeling the behavior of TCP congestion control. After trained with typical data samples, a three-layer (3-10-1) neural network model with fixed weights has been tested over a wide range of network conditions. In contrast to the equation models, our model can better associate the TCP factors, i.e. round trip time (RTT), retransmission timeout (RTO) and the loss event rate, with the throughput. Therefore, it can more accurately estimate the TCP throughput. As the estimation is done through the fixed neural model, the computational complexity is small, so it can be used for real-time online computing.