Entropy-based active learning for scheduling in wireless networks

It is well known that Max-Weight scheduling algorithm is throughput optimal, when the complete channel state information (CSI) is available at the scheduler. In this work, we address the joint design of scheduling and channel probing under general channel models. Our method predicts the instantaneous channel rates, and calculates the uncertainty in the prediction to make a scheduling and probing decision. To explicitly quantify the uncertainty in the channel prediction that will be removed by channel probing we adopt entropy measure from information theory. In order to accurately predict instantaneous user channel states we employ a Bayesian approach and use Gaussian processes as a state-of-the-art regression technique. We analytically prove that our algorithm achieves a fraction \epsilon of the full rate region when complete CSI is available. We demonstrate numerically under realistic assumptions that this rate region can be achieved by probing only less than 50% of all channels in a CDMA based cellular network utilizing high data rate protocol under practical channel conditions.