An Adaptive Congestion Control Protocol for Wireless Networks Using Deep Reinforcement Learning

Abstract

Congestion is more prevalent in wireless networks due to the unique challenges and limitations of the present wireless technology. With the increasing demand for high bandwidth and low latency communication, effective congestion control mechanisms have become crucial for ensuring efficient and fair utilization of network resources. Both the long and short flows of network traffic are the primary reasons for the congestion. However, the short flows cause a disproportionate amount of overhead and contribute significantly to congestion than long flows. Therefore, flow classification is imperative in congestion control design for the increased throughput. The existing conventional congestion control (delay and loss based) protocols do not adapt to current network scenarios on a timely basis and the current status that is learnt from the network environment is mostly not accurate. To overcome these limitations, machine learning techniques were used. However, due to real time constraints and dynamic network conditions, machine learning techniques may result in suboptimal performance. In this paper, a novel congestion control protocol is proposed especially for wireless networks using deep neural network (for flow classification) and deep reinforcement learning (for appropriate congestion window size). From the experimental results, it is evident that the proposed protocol improves the average network throughput against the state-of-the-art existing congestion control techniques such as TCP Cubic, TCP Ledbat, Bottleneck Bandwidth and Round-trip propagation time (BBR), and N-step temporal difference (N-step TD) learning technique by 25.6%, 23.46%, 21.85%, and 16% respectively.