Real-Time Distributed Coordination of Multi-Agent Systems under Limited Communication
Abstract
The field of multi-agent systems is growing rapidly thanks to the key role it plays in many militaryand commercial applications. One example is the autonomous operation of underwater networksfor oceanographic data collection, climate change monitoring, offshore exploration, pollutioncontrol and tracking, disaster prevention and tactical surveillance. Regardless of the specificapplication domain, the central goal is to deliberately and systematically shape the networkcollective behavior through the design of admissible local agent control algorithms. This isnontrivial due to various challenges, especially that i) tasks performed by the multi-agent systemalways vary with its environment, which is usually uncertain, time-varying, and potentiallyadversarial, ii) individual decision makers in a multi-agent system need to make local,computationally-friendly decisions in response to local observations and local communicationwith their neighbors, iii) communications between agents are never perfect due to packet loss, linkfailures, long latency, and/or limited bandwidth.The overarching goal of this proposal is to develop real-time, distributed decision-makingalgorithm to ensure high-performance and high-confidence operation of multi-agent systems inthe presence of time-varying uncertainty and limited communication, and to investigate theirfundamental performance limits due to various constraints in multi-agent coordination. To achievethis goal, the proposed research draws approaches from different disciplines such as onlineoptimization, dynamical programming/reinforcement learning, distributed optimization, statisticallearning, information theory, and control. The specific research thrusts entaili) design of online/real-time algorithms to handle time-varying uncertainty and to analyze onlineperformance such as the dynamic regret and competitive ratio. We will quantify how the onlineperformance changes with the volatility of the environment and the quality of future predictionsof uncertainties. Moreover, we will also investigate the inherent performance limit for broadclasses of online algorithms.ii) design of distributed online algorithms where agents make local decisions based on localobservations and local communication with neighbors. The distributed algorithms will achievenearly the same or comparable performance as the centralized algorithms. We will analyze theperformance using metrics such as the solution efficiency, optimality/dynamical regrets,constraint violation, convergence speed, and computation/communication overhead.iii)design of distributed algorithms with reduced communication using quantization as well asmaking the algorithms robust to communication failures and delays using tools such as passivityand the scatter transformation. We will analyze the communication bandwidth requirement byinvestigating the communication complexity of the quantized algorithms and also characterizethe tradeoff between communication, computation, and the solution efficiency.Besides the proposed theoretical and algorithmic research, we will also evaluate andexperimentally test the results on robotic swarm testbeds in Harvard, in particular, a kilobot swarmtestbed and an autonomous underwater mini-robot testbed. The theoretical development andexperimental testing will be performed in a close-loop fashion. Together, this proposed researchwill develop real-time distributed algorithms with limited communication that are adaptive to timevaryingenvironments, scalable to large scale systems, and robust to limited communication, thusadvancing the realization and autonomy of unmanned systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 24, 2019
- Source ID
- N000141912217
Entities
People
- Na Li
Organizations
- Office of Naval Research
- President and Fellows of Harvard College
- United States Navy