Distributed Optimization over Complex Networks
Abstract
This work focuses on developing theoretical insights for distributing optimization tasks among agents of multi-agent complex networks. In these setups, individual agents (network nodes) independently perform computation and communication tasks while all agents cooperate to solve network-wide optimization problems. These complex systems abound in practice (e.g., sensor networks and Unmanned Aerial Vehicles (UAVs)) and may be represented as static or dynamic networks, consist of sub-networks, and possibly include one or multiple super-agents (that control a set of system nodes-agents). To investigate the underlying hypotheses of utilizing distributed optimization in complex networks, we pose three fundamental questions; 1. What network attributes and problem characteristics justify solving an optimization problem in a distributed fashion over complex networks where centralized problem solving (through a super-agent that controls all agents) is accessible. To answer this question, we will -Establish a theoretically grounded connection between communication burden, network properties, and system-level performance of distributed problem-solving in complex networks. -Derive a mathematically rigorous relationship between optimization problem characteristics, agents’ computational responsibilities, and execution of networkwide distributed optimizations. -In networks that contain sub-networks (network-of-networks), what are the trade-offs between carrying out computations at the edge (agent level) versus distributing problem solving responsibilities between super-agents (each controlling a sub-network). The following steps will guide our answer to this question, -Develop convergence criteria to connect inner (intra-sub-network) and outer (inter-sub-network) communications rates with the network-wide performance of distributed optimization. -Study system-level performance of distributed optimization as a function of its sub-networks’ characteristics (e.g., sub-network size). 3. How can we extend our findings from static setups to dynamic networks where agents might exchange information with a different set of agents at any given time. Addressing this question, requires taking the below steps, -Extend the previously developed analysis to account for dynamic communication patterns. -Propose measures (e.g., best reinforcement options) to increase resiliency in contested environments and in the face of adversarial and non-adversarial disruptions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 29, 2024
- Source ID
- FA95502310203
Entities
People
- Javad Mohammadi
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Texas at Austin