Software and Demonstration

Software

• Code for our paper “A Deep Learning Framework for Optimization of MISO Downlink Beamforming" available at Github.

Demonstration

• Dynamic Channel Selection in Cognitive Radio Networks using the Multi-Arm Bandit Approach

Radio spectrum is a precious and expensive resource. Over years, we have seen exponential annual wireless data growth, so the demand keeps increasing for the finite resources of radio spectrum. To meet this demand, the current static and exclusive allocate spectrum is not effective, and sharing has become necessary.

In this demo, we will showcase the dynamic spectrum sharing, which forms part of the vision of cognitive radio. In such a system, different radio nodes will access the same frequencies and each node will use AI to determine how to optimize its parameters such as transmit power and frequency as the environment changes and share the spectrum with others dynamically.

The implemented Cognitive Radio (CR) system has been illustrated in Figure 1, where there is a pair of Second Users (SUs) want to communication using one of the four channels owned by a Primary User (PU). The demo illustrates a dynamic opportunistic channel access of SUs via CR, where the channel selection is made by the reinforcement learning algorithm called Multi-Arm Bandit (MAB) learned from both historical and instance channel sensing results.

Figure 1: The Cognitive Radio Demonstration System

During the experiment, the cooperative channel selection procedure of SUs is embedded with the data exchange procedure, i.e., the channel sensing result and the channel selection decision are exchanged between SUs via an extended ACK package for the data exchange procedure. This mitigates the overhead due to the exchange of sensing results and channel selection decisions during SU collaborations.

The channel selection decisions are learned using MAB based on both historical channel sensing results (illustrated as "Historical Channel Information" part in Figure 1), and the instant channel sensing result (illustrated as "Instant Channel Sensing Result" part in Figure 1). By learning from the channel quality (via energy detection in the demo), the implemented CR system can empirically select the channel which is not only less likely to be occupied by the PU, but also with the best predicted channel quality for data exchange. It achieves much higher data rate than both the random channel selection scheme and the energy detection algorithm as illustrated in Figure 2.

Figure 2: Performance comparison between our proposed algorithm and two benchmark algorithms: random channel selection and energy detection.

The channel selection decisions include two steps: the first is to select the next channel (illustrated as GREEN) for the data exchange (and CR collaborations since this is embedded in the data exchange), and the second is the backup channel selection (illustrated as YELLOW) if the next channel is occupied. Should the next channel for the data exchange be not available to use, i.e., occupied by PU, then the corresponding channel will be illustrated as RED and the SUs will automatically switch to backup channel. The SUs will then decide whether to stay in the backup channel or switch to a better channel. This avoids the interfere with PUs, as well as being interrupted by PUs.

A video is embedded for demonstration.

For more technical details, please refer to the paper below:

M. You, X. Zhang, G. Zheng, J. Jiang and H. Sun, "A Versatile Software Defined Smart Grid Testbed: Artificial Intelligence Enhanced Real-Time Co-Evaluation of ICT Systems and Power Systems," IEEE Access, vol. 8, pp. 88651-88663, 2020, doi: 10.1109/ACCESS.2020.2992906.