Dynamic ARF for throughput improvement in 802.11 WLAN via a machine-learning approach

The paper presents a dynamic Auto Rate Fallback (ARF) algorithm to improve the performance of aggregate throughput in IEEE 802.11 Wireless Local Area Network (WLAN). ARF is a simple and heuristic Rate Adaptation (RA) algorithm adopted by most of the commercial 802.11 WLAN products. However, when the traffic contentions among 802.11 nodes rise, using ARF will tend to degrade transmission rates due to increasing packet collisions and can consequently cause a decline of overall throughput. In this paper we propose a machine-learning based dynamic ARF scheme which utilizes neural networks to learn the correlation function of the optimal success and failure thresholds with respect to the corresponding contention situations including the number of contending nodes, channel conditions, and traffic intensity. At runtime, the generalized mapping function is then applied to determine the optimal threshold values depending on the current contention situations to achieve the best system throughput. We use the Qualnet simulator to evaluate and compare the performance of our scheme with that of the ARF and AARF algorithm. Simulation results illustrate that the proposed dynamic ARF approach outperforms these RA schemes in terms of improving the aggregate throughput in a variety of 802.11 WLAN environments.