CAREER: Scalable Architectures for Self-Managed Networks

该方案涉及基于个体用户强化学习的分布式网络体系结构。该计划的主要目标是为适用于大型网络的动态和分布式资源共享机制开发一个通用框架。考虑其中用户以若干可选方式之一访问网络资源的一般公式。每个用户被限制只能观察自己对结果的解释，而这个结果也取决于其他用户的行为。该框架涉及网络用户基于局部信息的非合作决策，进而在网络规模上是可扩展的。该研究计划将根据算法参数渐近精确的非线性微分方程来识别网络的宏观动力学。极限系统与进化生物学背景下出现的某些动力系统密切相关。成功完成的计划将确定可能的均衡制度，将确定导致稳定网络运行的分布式算法的特征，并将制定网络设计和管理指南，以维护网络所需的运行制度。我们将分析算法的稳健性，并通过模拟和实验验证所获得的结果。该计划的教育方面涉及课程开发和教学，最终目标是强大的研究计划和密切的产业合作，研究生和本科生的指导，以及本科生的积极研究参与。

The proposed program concerns distributed network architectures that are based on reinforcement learning by individual users. The main goal of the program is to develop a general framework for dynamic and distributed resource sharing mechanisms that are suitable for large networks. A generic formulation is considered in which users access network resources in one of several alternative ways. Each user is restricted to observe its own interpretation of the outcomes only, where this outcome depends also on the actions of other users. The proposed framework involves non-cooperative decision making by network users based on this local information, in turn it is scalable in the size of the network. The research program will identify macroscopic dynamics of the network in terms of nonlinear differential equations that are asymptotically exact in algorithmic parameters. The limit system is closely related to certain dynamical systems that arise in the context of evolutionary biology. Successfully completed program will identify possible equilibrium regimes, will characterize distributed algorithms that lead to stable network operation, and will develop network design and management guidelines that maintain desired operating regimes for the network. Robustness of the proposed algorithms will be investigated analytically, and verification of obtained results will be carried out via simulations and experiments.Educational aspects of the proposed program involves course development and teaching with the ultimate goal of a strong research program and close industrial collaboration, mentoring at both graduate and undergraduate levels, and active research participation from undergraduate students.