Optimal leader selection and demotion in leader-follower multi-agent systems

TL;DR

This paper develops a convex optimization framework for optimal leader selection and demotion in multi-agent systems, minimizing transfer function differences and providing explicit solutions for these problems.

Contribution

It introduces a convex relaxation approach to solve leader selection and demotion problems optimally, revealing that the relative H2 error depends only on leader counts.

Findings

Global optimal leader sets are identified via convex optimization.

Leader demotion solutions are characterized by fixed-size subsets.

Relative H2 error is independent of the total number of agents.

Abstract

We consider leader-follower multi-agent systems that have many leaders, defined on any connected weighted undirected graphs, and address the leader selection and demotion problems. The leader selection problem is formulated as a minimization problem for the $H^2$ norm of the difference between the transfer functions of the original and new agent systems, under the assumption that the leader agents to be demoted are fixed. The leader demotion problem is that of finding optimal leader agents to be demoted, and is formulated using the global optimal solution to the leader selection problem. We prove that a global optimal solution to the leader selection problem is the set of the original leader agents except for those that are demoted to followers. To this end, we relax the original problem into a differentiable problem. Then, by calculating the gradient and Hessian of the objective function of the relaxed problem, we prove that the function is convex. It is shown that zero points of the gradient are global optimal solutions to the leader selection problem, which is a finite combinatorial optimization problem. Furthermore, we prove that any set of leader agents to be demoted subject to a fixed number of elements is a solution to the leader demotion problem. By combining the solutions to the leader selection and demotion problems, we prove that if we choose new leader agents from the original ones except for those specified by the set of leader agents to be demoted, then the relative $H^2$ error between the transfer functions of the original and new agent systems is completely determined by the numbers of original leader agents and leader agents that are demoted to follower agents. That is, we reveal that the relative $H^2$ error does not depend on the number of agents on the graph. Finally, we verify the solutions using a simple example.