University of Maryland

Planning and Metareasoning for Multi-Agent Systems with Variable Communication Availability

This page contains the following sections:

Principal Investigators

Research Team

Tekevwe Akoroda, Akshay Bapat, Estefany Carrillo, Mohamed Khalid M Jaffar Sharan Nayak, Ruchir Patel, Eliot Rudnick-Cohen, Suyash Yeotikar.

Funding Source

This project is funded by the U.S. Air Force Research Laboratory.

Background and Project Objectives

The research is part of AFRL's Autonomous Swarms for Information-aware Mission Operations with Verification (ASIMOV) program, which studies algorithms for swarm autonomy in communications-constrained operating environments. The ASIMOV program is directed by Dr. Jeff Hudack at AFRL.

Coordinating the actions of different agents (such as unmanned aerial vehicles) to perform complex tasks like collaborative search and tracking, are essential to an autonomous swarm. The relative performance of coordination techniques may change as the availability of communication such as bandwidth and reliability varies.

This research project will develop hybrid centralized and decentralized multi-agent task allocation methods with different levels of communication availability. They also will build metareasoning techniques to enable autonomous agents to select the most appropriate task allocation method in real-time based on the current state of communication availability and other relevant factors. The team will evaluate the methods in situations that include a swarm of autonomous vehicles fighting a wildfire.


In this research project, we first evaluated the performance of different collaboration algorithms for decentralized task allocation in scenarios with different communication availability. In these scenarios, the agents in the multi-agent system (swarm) needed to find and then visit numerous targets in the region of interest. These were similar to collaborative search and tracking (CSAT) and wide-area search and surveillance (WASS) missions.

This video shows the performance of two algorithms (the asynchronous consensus-based bundle algorithm (ACBAA) and a decentralized genetic algorithm (GA)) that agents (here specified as UAVs) can use to coordinate their actions and complete the mission.

We then constructed a metareasoning policy; using this policy, each agent estimates the current communication availability and then selects the appropriate collaboration algorithm. The metareasoning policy is a rule that selects the communication algorithm based on the current estimate of communication availability.

We tested the metareasoning policy in scenarios in which communication availability changes, and the policy led to better system performance (reduced time to complete the mission).

Related Publications

Last updated on July 22, 2021, by Jeffrey W. Herrmann.