Multi-Agent, Multi-Scale Systems with the Koopman Operator

TL;DR

This paper extends the Koopman Operator framework to multi-agent systems with hierarchical and multi-scale dynamics, enabling analysis and control through centralized optimization and game theory.

Contribution

It develops a novel Koopman Operator formulation for multi-agent systems, incorporating hierarchical control and time-scale separation, and compares centralized and game-theoretic solutions.

Findings

Koopman-based approach facilitates multi-agent control analysis.

Coupling between agents affects social optimum and Nash equilibrium.

Hierarchical and multi-scale dynamics are effectively modeled with the new formulation.

Abstract

The Koopman Operator (KO) takes nonlinear state dynamics and ``lifts'' those dynamics to an infinite-dimensional functional space of observables in which those dynamics are linear. Computational applications typically use a finite-dimensional approximation to the KO. The KO can also be applied to controlled dynamical systems, and the linearity of the KO then facilitates analysis and control calculations. In principle, the potential benefits provided by the KO, and the way that it facilitates the use of game theory via its linearity, would suggest it as a powerful approach for dealing with multi-agent control problems. In practice, though, there has not been much work in this space: most multi-agent KO work has treated those agents as different components of a single system rather than as distinct decision-making entities. This paper develops a KO formulation for multi-agent systems that structures the interactions between decision-making agents and extends this formulation to systems in which the agents have hierarchical control structures and time scale separated dynamics. We solve the multi-agent control problem in both cases as both a centralized optimization and as a general-sum game theory problem. The comparison of the two sets of optimality conditions defining the control solutions illustrates how coupling between agents can create differences between the social optimum and the Nash equilibrium.