Game theoretic control of multi-agent systems: from centralised to distributed control

Abstract

Differential game theory provides a framework to study the dynamic strategic interactions between multiple decisors, or players, each with an individual criterion to optimise. Noting the analogy between the concepts of "players'' and "agents'', it seems apparent that this framework is well-suited for control of multi-agent systems (MAS).

Most of the existing results in the field of differential games assume that players have access to the full state of the system. This assumption, while holding reasonable in certain scenarios, does not apply in contexts where decisions are to be made by each individual agent based only on available local information. This poses a significant challenge in terms of the control design: distributed control laws, which take into account what information is available, are required. In the present work concepts borrowed from differential game theory and graph theory are exploited to formulate systematic frameworks for control of MAS, in a quest to shift the paradigm from centralised to distributed control.

We introduce some preliminaries on differential game theory and graph theory, the latter for modeling communication constraints between the agents. Motivated by the difficulties associated with obtaining exact Nash equilibrium solutions for nonzero-sum differential games, we consider three approximate Nash equilibrium concepts and provide different characterisations of these in terms a class of static optimisation problems often encountered in control theory. Considering the multi-agent collision avoidance problem, we present a game theoretic approach, based on a (centralised) hybrid controller implementation of the control strategies, capable of ensuring collision-free trajectories and global convergence of the error system.

We make a first step towards distributed control by introducing differential games with partial information, a framework for distributed control of MAS subject to local communication constraints, in which we assume that the agents share their control strategies with their neighbours. This assumption which, in the case of non-acyclic communication graphs, translates into the requirement of shared reasoning between groups of agents, is then relaxed through the introduction of a framework based on the concept of distributed differential games, i.e. a collection of multiple (fictitious) local differential games played by each individual agent in the MAS. Finally, we revisit the multi-agent collision avoidance problem in a distributed setting: considering time-varying communication graph topologies, which enable to model proximity-based communication constraints, we design differential games characterised by a Nash equilibrium solution which yields collision-free trajectories guaranteeing that all the agents reach their goal, provided no deadlocks occur. The efficacy of the game theoretic frameworks introduced in this thesis is demonstrated on several case studies of practical importance, related to robotic coordination and control of microgrids.