CIF: Small: Risk-Aware Resource Allocation for Robust Wireless Autonomy

无线自主网络系统（wan）几乎无处不在，执行各种通常复杂且普遍的以数据为中心的操作，例如传感、处理、学习和行动（即决策）。例子包括现代无线通信网络（例如，基于5G/6G，毫米波/太赫兹技术），无人机群，移动或机器人网络，无人驾驶飞行器（uav），自动驾驶汽车和物联网（IoT）。虽然广域网及其应用为社会和经济增长提供了巨大的潜力，但此类系统的运行不仅需要高效并由实际的、可观察的数据驱动，还需要满足通常严格的设计规范。这些规范是由维护性能健壮性和弹性的需要引起的，这反过来又转化为延迟、可靠性、公平性和可信赖性保证。这些标准和限制从根本上与无线系统操作相关的内在风险有关；这些风险是由于自然现象造成的固有不确定性造成的，例如无线媒体的非平凡统计分散以及多个用户和设备行为的随机性，通常具有复杂和异构的特征和目标。该项目提出了一种新的原则性方法框架，用于无线系统中系统风险感知资源分配，弥合了遍历风险中性和最小最大保守性之间的操作差距。研究重点不仅是新的约束风险意识资源分配问题的表述和双域变分分析，而且还包括最优风险意识策略的基于模型的综合和无模型强化学习的理论和有效方法的发展。期望这项工作将为无线系统资源分配建立一个新的范例。关于基本的程式化资源分配问题的初步结果——简单到单用户功率约束率最大化——表明风险意识策略相对于遍历策略（即风险中性策略）和极大极小策略都有明显的优势。然而，在更现实和有用的环境中获得最佳的风险意识策略是非常重要的：在风险意识问题中，期望的作用是由更一般的函数发挥的，称为风险度量，其中期望的基本性质-例如线性，同质性或塔性质-通常不存在。在受限的优化设置中，这种复杂性自然会被放大。本项目主要研究无线系统受限资源分配背景下的风险意识问题，分为三个重点：1)风险意识资源分配中的拉格朗日对偶性；2)基于模型的风险意识资源分配策略的数据驱动综合；3)风险意识资源分配策略的无模型学习。首席研究员预计，该项目将有助于改善风险意识设置下的固有挑战，例如基于风险度量的变分随机约束、资源策略的无限维度、随机服务的非凸性以及通道/系统模型可用性，最终使风险意识无线系统资源分配成为一项智力上可访问且计算上负担得起的任务。该项目还将涉及无线自治以外的几个领域，如金融、经济、能源和机器人，并可能引发通信、信息理论、统计和优化交叉领域的新发展，并激发风险意识和约束学习方面的新工具。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Wireless autonomous networked systems (WANS) are virtually everywhere around us, performing all kinds of often complex and pervasive data-centric manipulations, such as sensing, processing, learning, and acting (i.e., decision-making). Examples include modern wireless communication networks (e.g., based on 5G/6G, mmWave/THz technologies), drone swarms, mobile or robotic networks, unmanned aerial vehicles (UAVs), self-driving cars, and the Internet of Things (IoT). While WANS and their applications present a high potential for societal and economic growth, the operation of such systems requires not only to be efficient and driven by actual, observable data but also to meet often strict design specifications. These specifications are induced by the need to maintain performance robustness and resilience, which, in turn, translates into latency, reliability, fairness, and trustworthiness guarantees. Such criteria and constraints are fundamentally connected to intrinsic risks associated with the operation of wireless systems; those risks are due to inherent uncertainties caused by naturally occurring phenomena such as nontrivial statistical dispersion of the wireless medium as well as randomness in the behavior of multiple users and devices, often with complex and heterogeneous features and objectives. This project puts forward a new principled methodological framework for systematic risk-aware resource allocation in wireless systems, bridging the operational gap between ergodic risk neutrality and minimax conservativeness. The investigation focuses not only on formulation and dual-domain variational analysis of new constrained risk-aware resource-allocation problems but also on the development of theory as well as efficient methods for both model-based synthesis and model-free reinforcement learning of optimal risk-aware policies. It is expected that this work will establish a new paradigm in wireless systems resource allocation.Preliminary results on basic stylized resource-allocation problems - as simple as single-user power-constrained rate maximization - demonstrate clear advantages of risk-aware policies against both their ergodic (i.e., risk-neutral) and minimax counterparts. However, obtaining optimal risk-aware policies in more realistic and useful settings is nontrivial: in risk-aware problems, the role of expectations is played by more general functionals, called risk measures, for which fundamental properties of expectation - such as linearity, homogeneity, or the tower property - are generally absent. Such complications are naturally amplified within a constrained optimization setup. This project concentrates on such risk-aware problems within the context of constrained resource allocation for wireless systems and is divided into three main thrusts: 1) Lagrangian duality in risk-aware resource allocation, 2) model-based data-driven synthesis of risk-aware resource-allocation policies and 3) model-free learning of risk-aware resource-allocation policies. The principal investigator anticipates that the project will be instrumental in ameliorating inherent challenges under the risk-aware setting, such as the presence of risk-measure-based variational stochastic constraints, infinite dimensionality of resource policies, nonconvexity of random services, and channel/system-model availability, ultimately rendering risk-aware wireless-system resource allocation an intellectually accessible and computationally affordable task. The project will also be relevant to several areas beyond wireless autonomy - such as finance, economics, energy, and robotics - and may trigger new developments in the intersection of communications, information theory, statistics, and optimization, as well as inspire new tools in risk-aware and constrained learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.