ReLoC - Reinforcing Lowered Congestion

Jensen, Jakob Lund ; Valentinussen, Peter Rudbæk

4. semester

Computer Engineering, Master

2024

2024-05-31

115 pages

This project sets out to explore the possibilities of applying deep reinforcement learning to TCP congestion control. Throughout the project, a novel deep reinforcement learning-based congestion control algorithm was designed and implemented called Reinforcing Lower Congestion (ReLoC). Furthermore, three different variations of ReLoC were created and tested in a variety of network scenarios. Four policies were created for the three variations of ReLoC. These four policies were the results of training in four different environments. The first policy was trained using a network environment that would always follow the same pattern of changing available capacity without any other TCP clients communicating over the channel. The second policy was trained using a network environment that would switch to random levels of available capacity, but still without any other TCP clients communicating over the channel. The final two policies were trained using the same network environments as the two prior respectively, but with another TCP client communicating over the channel. Through testing it was shown that the ReLoC variants were able to outperform commonly used congestion control algorithms when operating alone on the channel, however, it would underperform when another TCP client would be utilizing the channel as well. Various suggestions for further improvements have been presented which may help ReLoC being able to perform well in scenarios where it is sharing the channel with another TCP client.

Download
View record in AAU Student Projects