Automatic MAC protocol selection in wireless networks based on reinforcement learning.

Existing MAC protocols do not address the dynamism and complexity of the environment and applications of today’s wireless networks. The running applications and the environment change all the time, and as a consequence the requirements of the wireless transmissions change. For example, in one moment a client is transmitting video, requiring high throughput; next it will control a robotic arm, requiring bounded delays. In this example, a contention-based MAC protocol would cope with flexible traffic demands, however it does not meet the delay constraints. Reservation based protocols, meanwhile, provide performance guarantees, but at a higher overhead. Hence, wireless networks require adaptive techniques that change how the network reacts over time. To that end, we propose SOMAC (Self-Organizing MAC), a system that uses reinforcement learning techniques to switch the MAC protocol in structured wireless networks according to the ongoing network demand. The novelty of SOMAC lies in its use of reinforcement learning, which solves the following shortcomings in the literature: (i) the lack of models that cope with changes in its environment or lack of representative data during training; (ii) the capacity to self-optimize based on a number of metrics. To showcase its genericity, we evaluated the model using two different optimization metrics (throughput and delay) on a testbed.