EAGER- DynamicData: Novel Approaches for Optimization, Control, and Learning in Distributed Networks

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Recent advances in sensor and robotics technology have led to large networks of coordinated mobile platforms that can perform automatic sensing, mapping, learning, and control tasks. While the remote tasks can be controlled by the ground center via long-range communication links, the links are expensive and suffer long delays. This project develops computational methods that will significant improve the ability for the remote agents to coordinate locally with one another for tasks such as recognize and navigate around obstacles, reconstruct signals, and learn new control policies from a large amount of multi-modal sensor data, all done with little or no communication to a center. The proposed approach is radically different from the current state of the art. It also includes educational components such as courses, seminars, and initiatives for under-represented minority and women. The proposed work is a set of novel algorithms for a variety of computing problems in multi-agent networks, for problems involving extremely large-scale distributed datasets and high complexity objectives, and with greater accuracy at rates that are provably faster than existing methods. Very promising preliminary results have been obtained. The proposed project includes (a) a new approach to integrate multi-agent coordination with problems arising in optimization, game theory, control, and learning, into systems of equations, inclusions, or variational inequalities; (b) novel operator splitting methods that lead to decentralized numerical solutions of these systems, which scale to new levels of size, complexity, and diversity; (c) stochastic approximation techniques to deal with the imminent "distributed data deluge", along with accelerations techniques based on variance reduction, importance sampling, and asynchronous parallelization; and (d) a set of open-source software products for optimization, control, and learning problems with dynamic and large-scale data, along with a comprehensive evaluation plan. The contributions of the project is a unified framework in parts (a) and (b) above, which enable the decentralized numerical solutions at new levels of speed, complexity, diversity, and resilience. In order to achieve the goals, substantial resources will be devoted to both mathematical research and engineering challenges.