NeTS:Small:Optimal Learning Times for Task-Oriented Communication Networks

通信网络必须支持一组不同的任务，同时快速适应时变的条件。该项目致力于开发和表征快速、自适应的网络控制方法。这些方法尊重网络的能量、计算、通信和感知资源，同时保持每个任务的高质量执行，并在所有用户之间提供公平的共享。我们的目标是创建更智能的网络，具有更快的响应时间和更低的延迟，同时适应用户移动性和设备功能的变化。该研究在控制理论、运筹学、智能电网调度、经济学、博弈论等领域有着广泛的应用。这个项目还将在这些主题上培训几名研究生。这项工作具有挑战性，因为每个任务的时间变化、流动性和不同步的开始和停止时间。基本的优化问题涉及时间平均的非凸比率，这些比率还没有被显著地研究过。最优收敛时间是未知的。主要研究人员的前期工作提出了一种改进次梯度凸规划收敛时间的新方法。这个项目试图扩展这种方法，并开发新的方法来解决面向任务的随机网络的更困难的问题。这项工作的另一个组成部分是在这种情况下制定收敛时间的基本下界。这可能会对优化、随机控制、在线决策和机器学习等更广泛的领域产生很大影响。

Communication networks must support a diverse set of tasks while quickly adapting to time-varying conditions. This project seeks to develop and characterize fast and adaptive network control methods. The methods respect the energy, computation, communication, and sensing resources of the network while maintaining high quality execution of each task and providing fair sharing across all users. The goal is to create smarter networks with faster response times and lower delays while adapting to changes in user mobility and device functionality. The research has broad applications to control theory, operations research, smart grid scheduling, economics, and game theory. This project will also train several graduate students on these topics.This work is challenging because of time variation, mobility, and asynchronous start and stop times of each task. The fundamental optimization problems involve nonconvex ratios of time averages that have not been significantly explored. The optimal convergence times are unknown. Preliminary work of the principal investigator develops a new method for improving convergence time of subgradient-based convex programming. This project seeks to extend this method and to develop new methods for solving the more difficult problems of task-oriented stochastic networking. Another component of this work is the development of fundamental lower bounds on convergence times in this context. This can have a high impact on broader areas of optimization, stochastic control, online decision making, and machine learning.