Machine learning for congestion control

Abstract In this chapter, we discuss a sender-side congestion control algorithm based on deep reinforcement learning. After a review of network congestion, methods for avoidance are discussed, including the algorithm of this chapter. The algorithm learns to correlate congestion signals with the value of rate control decisions to maximize its own expected reward. Its objective is to converge to a sending rate that yields a throughput equivalent to the capacity of a transmission path without generating a standing queue at the bottleneck. The chapter provides several examples implemented for the ComNetsEmu as Python scripts and concludes with hands-on examples the reader can perform beyond the examples presented here.