Controlling A Large Network of Inexpensive Agents

There has been tremendous advancements in high performance robots,
drones, and other autonomous or semi-autonomous systems. At the same
time, interest is growing in use of large networks of (relatively)
inexpensive agents for a variety of tasks.  This often implies
limitations on power (actuation), bandwidth (communication frequency),
homogeneity of agents, or configuration reliability (loss or addition of
some agents during operation).  Strong results exist for most of these
challenges at a single agent level and they all result some level of
conservatism (degradation of performance guarantees), which are more or
less inevitable.

Extending these techniques to multi-agent systems leads to an additional
source of conservatism: the structure of the network, as captured by the
Adjacency or Laplacian matrices.  The goal is to develop an approach for
agents to optimize their network weights for improved performance, in
real time, autonomously (decentrally), and with low computational
sophistication. The proposed solution is a mixture of known  -- in
original form or somewhat modified - techniques: max consensus,
augmented Lagrangian, power iteration, etc.