Many novel DoD missions, from disaster relief to cyber reconnaissance, require teams of humans and
machines with diverse capabilities. Current solutions do not account for heterogeneity of agent capabilities,
uncertainty of team knowledge, and dynamics of and dependencies between tasks and agent roles, resulting in
brittle teams. Most importantly, the state-of-the-art team design solutions are either centralized, imposing role
and relation assignment onto agents, or completely distributed, suitable for only homogeneous organizations
such as swarms. Centralized design models can’t provide insights for team’s self-organization, i.e. adapting
team structure over time in distributed collaborative manner by team members with diverse expertise and
responsibilities.
In this paper we present an information-theoretic formalization of team composition and structure adaptation
using a minimization of variational free energy. The structure adaptation is obtained in an iterative distributed
and collaborative manner without the need for centralized control. We show that our model is lightweight,
predictive, and produces team structures that theoretically approximate an optimal policy for team adaptation.
Our model also provides a unique coupling between the structure and action policy, and captures three
essential processes of learning, perception, and control.
|