Classical transportation systems have been suffering from inefficiency in utilizing road capacity and incapability of guaranteeing safety due to inconsistent and unpredictable human based manipulations. This long standing issue is expected to be resolved through the introduction of vehicular autonomy, along with a properly configured traffic management system and an underlying communication network that supports the dissemination of safety and control messages. However, mechanisms to be used in the regulation of autonomous vehicles are yet to be developed.

For the first part of the dissertation, we propose an efficient data dissemination protocol, identified as Vehicular Backbone Network (VBN), which support high throughput and low delay message delivery across a Vehicular Ad Hoc Network (VANET). Vehicles are elected as relay nodes at properly selected nominal positions to optimize system throughput rate with guaranteed outage probability. The proposed VBN mechanism is analyzed under different Medium Access Control (MAC) protocols and various stochastic traffic flow conditions.

In the second part of the dissertation, we go beyond the statistical vehicular traffic model

and present our study of a platoon-based autonomous vehicular traffic regulation system. The traffic management system, serving to configure vehicles into platoons and regulating the on-ramp access of vehicular flows into the highway, is integrated with a design of a VANET system. The latter provides for the communications of essential message flows among highway vehicles. The fundamental tradeoffs that need to be made in designing the autonomous systems are modeled and studied. We propose an optimization framework to be used in designing the traffic control system in a manner that meet throughput and latency performance metrics across the vehicular mobility and the message communications networking dimensions of the overall system. We also study the design of a core network that employs Road Side Units (RSUs) that provide communications networking support in disseminating message flows to/from and among highway vehicles.

Aging and inflammation have been shown to reduce the supply of functional B-cells; the prevailing explanation is that hematopoietic differentiation decisions are skewed towards the myeloid lineage. Here, we have addressed how inflammaging may affect B-lymphopoiesis itself. Using a new NFκB reporter mouse, we uncovered dramatic dynamics in RelA control during early B lymphopoiesis in young but not in all aged mice. When genetically perturbing NFκB RelA dynamics with specific IκB mutants, we found severe, cell-intrinsic defects in B lymphopoiesis, that could not be attributed to a differentiation block or reduced cell survival. Careful quantification of B-lymphoid progenitors allowed us to fit a mathematical model of the differentiation pathway, which led to the seemingly paradoxical prediction – confirmed ex vivo – that mutant B-cell progenitors ‘rush through’ the differentiation pathway. Further analysis revealed the differentiation hypermorph leads to diminished mature cell populations given that differentiation pauses are essential for population expansion. Transcriptomic profiling at single-cell resolution revealed RelA dynamic dysregulation resulted in an accelerated progression of transcriptomic cell states that are not in sync with the classically defined progenitors and – importantly – overridden preBCR checkpoint. As this is a major quality control step to ensure proper immunoglobulin heavy chain recombination, the mutant triggers premature κ-light chain recombination, resulting in B-cells with nonfunctional immunoglobulin. Our findings establish a new paradigm for how aging and inflammation impact humoral immunity via dysregulation of RelA dynamics: accelerated differentiation diminishes the size and the quality of the B-cell pool available for adaptive immune responses.

Massive MIMO technology, leveraging beamforming and precoding, significantly enhances spectrum and energy efficiency when CSI at the transmitter is available. However, the large-scale arrays envisioned for millimeter-wave or terahertz communications drastically increase the downlink CSI training overhead and uplink feedback overhead, necessitating efficient CSI feedback designs in FDD wireless systems. Similar to image compression, after downlink CSI estimation at the UE, downlink CSI is encoded into a low-dimensional codeword stream, transmitted to the base station, and decoded for downlink CSI recovery. Recently, deep learning, particularly autoencoders, has been widely discussed as an efficient CSI feedback approach, outperforming traditional compressive sensing methods. However, several challenges remain unaddressed in learning-based CSI feedback frameworks.

In this dissertation, we address six critical issues in learning-based CSI feedback frameworks. We explore the exploitation of uplink/downlink frequency-division duplexing reciprocity. The energy, delay, and angles of arrival and departure of uplink and downlink channels are highly correlated. Yet, uplink CSI, available at the base station, is seldom used to reduce the uncertainty of downlink CSI recovery. We propose a deep learning CSI feedback framework leveraging this reciprocity, with a redesigned loss function for joint encoding of CSI magnitudes and phases. Evaluation shows superior performance and better utilization of frequency-division duplexing reciprocity compared to previous works.

We also address the reduction of pilot transmission overhead. Given limited pilot resources, it is impractical to transmit pilots for all antenna ports in a massive MIMO system. We introduce a beam-based pilot precoding approach and a deep learning CSI feedback framework to minimize pilot transmission and CSI feedback overhead from UE. Furthermore, we propose scalable CSI encoding. Existing frameworks often encode and decode full CSI using autoencoders, leading to heavy models and low scalability. Recognizing low correlation between widely spaced antennas, we propose a scalable deep learning CSI feedback framework using a divide-and-conquer principle to encode CSI subarray-by-subarray. This dynamic compression approach achieves significant model size reduction while maintaining downlink CSI recovery performance.

To tackle the reduction of training costs, deep learning models typically require extensive data collection and customization for different channel types. Inspired by the simplicity and generality of JPEG in image compression, we propose a JPEG-based CSI feedback approach, which requires no prior training and adapts to various channels, offering comparable recovery performance to learning-based methods. Additionally, we enhance frequency selective channel performance. Current learning-based models struggle with CSI recovery in frequency selective channels due to sparse pilot placement. We propose an uplink CSI-assisted CSI upsampling module at the base station, compatible with most previous explicit CSI feedback frameworks. Ensuring adherence to standardized feedback, we note that current learning-based CSI feedback methods do not strictly follow standardized CSI feedback, making industry adoption challenging. We propose a lightweight, efficient precoder upsampler as a plug-in module for the base station, enhancing performance in high delay-spread channels.

For prospective researchers, we suggest two directions to bridge artificial intelligence research and cellular communications industries. Addressing channel aging in CSI feedback is crucial due to the time lag between downlink CSI training and downlink data transmission, which may result in outdated precoders. Future research should focus on developing learning-based CSI and precoder prediction to adapt to non-linear channel variations. Additionally, in practical frequency-division duplexing systems, the base station lacks exact CSI knowledge from UE and relies on fed-back precoders. Researchers should develop precoder-based user scheduling algorithms to avoid interference and maximize system throughput.

The building industry is responsible for over 40% of the world's energy consumption, making it imperative to develop more efficient building designs to reduce energy usage and minimize environmental impact. However, the current building design process is highly fragmented, resulting in inefficient and sub-optimal designs. Although the digitalization of the building industry happened nearly at the same time as the electronics industry, the latter has become fully automated, whereas the building industry still lags behind.

This dissertation identifies the challenges in the building design process that hinder automation and proposes a platform-based approach as part of the solution. Platform-based design enables the development of common interfaces for greater consistency and efficiency in the design process, along with modularity that breaks down complex systems into smaller, more manageable modules. This facilitates easier design iteration and adaptation to changing requirements, while also allowing for the reuse of design components across different projects, saving time and resources.

In addition, the dissertation explores the potential of levels of abstraction in building energy models to better support the design process without requiring a full building model at early design stages. This can help reduce design iterations and save time and costs. Finally, we streamline the process of constructing building models as a step towards automated design. Although fully automated design is yet to be realized, this streamlined process serves as a potential starting point for achieving automation in the future.