Intelligent-CW: AI-based Framework for Controlling Contention Window in WLANs

Abstract

The heterogeneity of technologies that operate over the unlicensed 5 GHz spectrum, such as LTE-Licensed-Assisted-Access (LAA), 5G New Radio Unlicensed (NR-U), and Wi-Fi, calls for more intelligent and efficient techniques to coordinate channel access beyond what current standards offer. Wi-Fi standards require nodes to adopt a fixed value for the minimum contention window (CW min ), which prohibits a node from reacting to aggressive nodes that set their CW min to small values. To address this problem, we propose a framework called Intelligent-CW (ICW) that allows nodes to adapt their CW min values based on observed transmissions, ensuring they receive their fair share of the channel airtime. The CW min value at a node is set based on a random forest, a machine learning model that includes a large number of decision trees. We train the random forest in a supervised manner over a large number of WLAN scenarios, including different misbehaving and aggressive scenarios. Under aggressive scenarios, our simulation results reveal that ICW provides nodes with higher throughput (153.9% gain) and 64% lower frame latency than standard techniques. In order to measure the fairness contribution of individual nodes, we introduce a new fairness metric. Based on this metric, ICW is shown to provide 10.89× improvement in fairness in aggressive scenarios compared to standard techniques.

Adaptive Contention Window Design Using Deep Q-Learning

“…Classical Q-learning performs well in settings where the stateaction space is small, since the table-like iterative procedure in (3) updates each of the |S| × |A| pairs separately. By contrast, in our setting we have an agent (node 0) whose action space can potentially be large (different choices for the MCW) and, more critically, the observations o k from the environment are real-valued.…”

“…Moreover, the number of wireless devices is constantly increasing, thus precipitating the urgent need for smart and efficient access to the limited available spectrum. Though oftentimes preferred due to their simplicity, conventional random access protocols such as IEEE 802.11 lead to suboptimal spectrum utilization and, more importantly, the performance of these protocols drops significantly under dynamic and uncertain settings [1][2][3].…”

“…The proposed uncertain scenario is of empirical relevance in cases where nodes may act unfairly by reducing their MCW to acquire more channel access at the expense of others. In this setting, we would rely on intelligent agents to adapt their own MCW to restore fairness [3]. A similar situation might arise when a new node joins a network and, without full knowledge of others' parameters, the new node seeks to learn their dynamics and predict the optimal MCW value at every time.…”

“…Motivated by the recent success of machine learning applied to wireless networks [17][18][19][20][21][22], learning-based solutions have been proposed for optimal distributed CW design [3,[23][24][25][26]. Germane to our approach is [24], where (non-deep) Q-learning is applied.…”

“…However, since the Q-function is learned as a table instead of parametrized by a neural network, the authors deviate from the current protocol and consider a different (much smaller) state-action space to enable learning. Another relevant approach is presented in [3], which considers a similar setting to ours but proposes a supervised learning strategy based on a random forest algorithm. As such, it performs well in a static scenario but fails to generalize to the cases where other nodes vary their CW parameters, as we illustrate through experiments.…”