Optimizing network control and resource allocation in large scale, ultradense millimeter-wave networks

Abstract

Wireless communication is a transformative technology that has changed the way we work, communicate and enjoy our free time. The number of connected devices is expected to increase to 29.4 billion by 2030. As a result, there is a demand for higher data rates, lower delays and constant connectivity of wireless devices. These demands have driven the networking community to seek new technologies as these requirements are beyond the capabilities of current networks. Communication at higher frequencies, beyond 10 GHz where most of current communications systems operate, could be the game changing technology for the next generation of networks. At millimeter-wave frequencies, 30-300 GHz, the available spectrum is larger than all spectrum currently allocated to cellular and wireless area networks (WLAN). The unlicensed spectrum at 60 GHz alone can offer 10 to 100 times more spectrum than it is available in current unlicensed WLANs. The larger bandwidth allocations allow for increased datarates. These multi-gigabit per second rates and milisecond latencies are now achievable thanks to new directional high gain antennas and cost-effective CMOS technology that can operate at mm-wave frequencies. However, with the use of this new technology new challenges arise. For example, since the communication is directional in mm-wave, the best direction to communicate needs to be decided, and updated in mobility scenarios. Aside from the technical deployment challenges at this frequency, another challenge is how to provide increased capabilities for the network such as localization and passive localization. The specific characteristic of propagation at mm-wave frequencies such as the quasi optical patterns, the dense deployments and the multi path diversity warrant designs of localization schemes aimed at exploiting these characteristic. Finally, stepping away from conventional network design, the future generation of networks can be enhanced by Machine Learning. Machine Learning has shown to be a powerful tool in many domains and has increased the autonomy of chatbots and self-driving cars. As we go further in frequency, ML can help design the different layers of the future communication networks to embrace the hardware imperfections, that are more relevant at higher frequencies, and to serve applications in the optimum way in different environments. The thesis is divided into two parts, first we study the challenges in the optimization and features of conventional mm-wave networks, and second, we consider a new way of designing mm-wave communications using Machine Learning. In the first part, we study the challenge of scaling mm-wave networks to dense deployments, and secondly we increase the capabilities of mm-wave networks by proposing a passive localization mechanism that is standard compliant. Currently the beam training mechanisms deployed to establish a directional link result in a large overhead in communications and this prohibits dense deployments. This is specially due to station beam training. While the APs can beam train with multiple stations simultaneously, the stations have to train one at a time. This overhead is even larger in mobile deployments where the devices require constant re-training to maintain communication. To overcome this overhead and allow dense deployments we propose a low-overhead beam training mechanism, SPIDER, that eliminates the station beam training overhead entirely. To this end, stations carry out phase-coherent measurements by switching through multiple receive beam patterns on a time-scale of tens of nanoseconds when receiving a packet preamble to calculate the AoA to the AP and select the best beam pattern for communication. Our results show that our algorithm achieves highly accurate angle estimation used to drive the beam steering decisions and that it reduces the overhead by an order of magnitude compared to IEEE 802.11ad beam training. The second work presented in the first part of this thesis is regarding the use of the network as a sensor in mm-wave frequencies. Joint communication and sensing approaches would allow for an extremely distributed cost-effective sensor network. At higher frequencies, due to the larger bandwidth, the theoretical precision of localization information is higher than at lower frequencies. However, localization algorithms for mm-wave frequencies can not directly build from the lower frequency ones. At lower frequencies, these algorithms leverage information from multiple antennas with separated RF chains, while in mm-wave phased antenna arrays are usually connected to a single RF chain, so the phase information per antenna can not be obtained. However, several properties of mm-wave communications have implications in the design of localization systems. The high path loss attenuation due to the high frequency implies that the line of sight paths (LOS) can be distinguished from the NLOS paths. Additionally, reflections on surfaces have limited scattering. These properties generate channels with sparse multi paths, for which geometrical propagation assumptions become appropriate. Additionally, the need for directional communication due to the high path loss results in the need for mechanisms that enable beam selection, which are the focus of the earlier works. Such beam training mechanisms become relevant for localization in mm-wave frequencies. Since during the beam training the environment is scanned using beam patterns in different directions, these information can be used to detect and track moving reflections. We present POLAR, the first passive mm-wave localization algorithm for mm-wave commercial devices with zero overhead. We use commercial APs whereas the station design is based on a full-bandwidth 802.11ad compatible FPGA-based platform with a phased antenna array. The stations exploit the preamble of the beam training packets of the APs to obtain channel state information measurements for all antenna patterns. With this, we determine distance and angle information for the different multipath components in the environment to passively localize a mobile object. We evaluate our system with multiple APs and a moving robot with a metallic surface. Our results show that is is possible to obtain 6.5cm mean error accuracy and sub-meter accuracy in 100% of the cases performing joint localization and sensing in a mm-wave testbed. In the second part of this thesis, we study a new way to design communications with the concept of AI-native air-interface framework in mm-wave. In this area of research, some works have proposed to learn to communicate using an end-to-end communication approach using deep learning (DL). However, a channel model is needed in order to train these models, and analytical channels were shown to produce a reduced performance in experimental scenarios. In order to extend this to real channels, model-free approaches have been presented in which the model can be learned implicitly from data. Building on this works to obtain an explicit channel model would also allow for transfer learning to speed up the training on time-variant channels. In this thesis, we present a neural network architecture based on Mixture Density Networks to explicitly learn the channel model for end-to-end communications for SISO and MIMO systems. Our results show that it is possible to learn complex stochastic channel models from data with differentiable methods that are compatible with the training of the autoencoder structure. These end-toend learning methods will become increasingly important in future generations of wireless networks, where the system complexity will increase due to a higher number of antennas, higher frequencies, wider bandwidths, and more complex hardware, making ML-based physical layer design highly appealing. Second, we study the scaling properties of end-to-end ML models for large-scale MIMO systems. We conclude that the bottleneck that prevents the scaling of such networks, in the state-of-the-art architectures, is due to the input and output representations. Large MIMO systems constitute highly demanding scenarios in which training would require unmanageable memory resources. We study typical approaches available in DL to reduce the memory allocation and show that a bit-wise architecture for MIMO allows to reduce the memory allocation by several orders of magnitude. Our results show that with the bit-wise architecture it is possible to extend the end-to-end approaches to large systems of antennas. This study is an important first step towards developing practical end-to-end learning algorithms for large scale MIMO systems. So far, end-to-end systems have only been applied in a per symbol manner as training these systems for blocks of symbols is complex as it introduces implicit channel coding. Optimizing the receiver only would allow to reduce the complexity and train in blocks of symbols while helping to deal with imperfections and other non-linearities without having to explicitly correct them. We explore the gains that can be achieved by using an ML based design on the receiver of single carrier mm-wave systems for frequency selective channels, where due to the multi-path and the resulting inter symbol interference block wise treatment is required. Our results show that it is possible to outperform baseline typical receivers in static scenarios. In summary, mm-wave technologies will open the door to the next generation of applications, however, new challenges need to be tackled. We provide new solutions to enable scalable deployments and show that the beam training station overhead can be completely eliminated. We introduce the feature of passive localization and sensing for mm-wave frequencies, and show the high accuracy achievable in an experimental FPGA based testbed. Finally, explore new ways of designing the physical layer to reduce the effect of hardware imperfections in mm-wave and show gains compared to state-of-the-art receivers.