CCSS: 2-Stage Stochastic Wireless System Optimization In the Field

About This Grant

Design/resource variables in wireless systems come in two special flavors: Short-term, dealing with dynamic and immediate decision-making to meet current demands or optimize instantaneous utility, and long-term, focusing on strategic planning over longer time horizons, optimizing a given wireless system in a certain statistical sense. As a matter of fact, short/long-term decision variables most frequently coexist in wireless system design, ultimately asking for joint management and distribution of resources as standard as bandwidth, power, phase and spectrum over different time scales (or stages), so as to optimize performance relative to various criteria, and on the basis of information with varying levels of detail. In this project, the PI puts forward a new methodological framework for the systematic treatment of joint short/long-term stochastic wireless system optimization. Such problems are formulated as natural instances of 2-stage stochastic programming with continuous uncertainty, and are considered under a fully data-driven and minimally model-free setting, allowing for (hopefully) deployable optimization schemes for training optimal wireless systems in-the-field. Concurrently, the proposed research also addresses several open technical challenges in algorithm development for solving general 2-stage stochastic programming models. Therefore, the proposed research is expected to be of more general interest and generate broader impact beyond the area of wireless systems as well. The proposed research is divided in the following two major thrusts: 1) 2-Stage wireless system design under imperfect oracles, addressing fundamental challenges, namely, establishing stochastic (sub)gradient representations of optimal values of (second-stage, short-term) problems with complex (i.e., non-trivial) constraints, dealing with inexactness caused by nonconvex optimization solvers, and planning against imperfect Channel State Information (CSI) estimation at the absence of ground truth information. In sharp contrast to existing work, the PI's goal here is to achieve fully data-driven, model-free system optimization, under no assumptions on (expensive to fit and maintain) cascaded CSI models and network structure. 2) Policy Function Approximations (PFAs) for model-free 2-stage system design, exploring functional reductions of 2-stage stochastic programs for eliminating instantaneous second-stage optimization, and enabling joint short/long-term training realizable by leveraging finite-dimensional policy parameterizations (i.e., PFAs) for short-term system optimization. The PI will investigate the design of PFAs with trainable constraint embeddings, as well as model-free synthesis of optimal policies, i.e., domain-inspired PFA discovery and training via zeroth-order stochastic optimization. Lastly, the PI will exploit such PFAs for learning distributionally robust short-term resource allocations with individual robustness levels (e.g., at users/terminals), targeting system effectiveness and deployability. 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.