Massive mimo: random access protocols based on statistical inference and machine learning techniques

Abstract

The number of fifth-generation (5G) cellular network devices is growing at an unprecedented rate. 5G technology is characterized by the ability to provide three types of essential services: enhanced Mobile Broadband (eMBB), Ultra-reliable Low Latency Communication (URLLC), and massive MTC (mMTC). These services support many types of applications such as virtual reality, augmented reality, traffic control, Internet of Things (IoT), industrial IoT, and others. To meet these demands, several technologies to support 5G and beyond (B5G) have been developed in recent years. Among these technologies are Intelligent Reflecting Surfaces (IRS), non-orthogonal multiple access (NOMA), millimeter-wave (mmWave), and massive multiple-input multipleoutput (MIMO). Of all these technologies, massive MIMO is the most successful. Massive MIMO is a major enabler for the implementation of mMTC technology, whose devices will be available in massive numbers and will require low power consumption and high connectivity. However, since the time and frequency resources provided by Base stations (BSs) are scarce, and the number of devices keeps increasing, it is likely that in the future, there will be a lack of pilot signals to serve all devices in the network, leading to a performance bottleneck in the system. To solve this problem, some random access protocols have been developed. This is the case of the strongest-user collision resolution (SUCRe) protocol, a grant-based (GB) protocol that grants access to network resources only to users that are the strongest contender for a particular pilot. Other protocols of the grant-free (GF) type, which is more suitable for supporting mMTC systems, are also proposed. In this work, three random access protocols are proposed. The first one is based on the SUCRe protocol and aims to optimize the decision step of the SUCRe protocol through a Bayesian classifier. This first protocol shows better results than SUCRe, and, interestingly, the effect of using the Bayesian Classifier (BC) only changes the decision criteria of the protocol to a more effective one. The second protocol is similar to the first but uses a classifier based on a Multilayer perceptron (MLP) Neural Network (NN) instead of a Bayesian classifier, presenting slightly better results. The third protocol is a GF Random-Access (RA) procedure and uses a reinforcement learning algorithm called Q-Learning to guide devices toward pilot signals that are less congested. The congestion levels are obtained in a massive MIMO simulation scenario with the realistic propagation effects considered. The algorithm stood out compared to traditional methods and comparison references, showing better results in metrics such as network throughput, per-user throughput, and latency. The three protocols presented also showed robustness in relation to variations of some parameters, reinforcing their efficacy.