Density Control of Multi-Agent Systems with Safety Constraints: A Markov Chain Approach

Abstract

The control of systems with autonomous mobile agents has been a point of interest recently, with many applications like surveillance, coverage, searching over an area with probabilistic target locations or exploring an area. In all of these applications, the main goal of the swarm is to distribute itself over an operational space to achieve mission objectives specified by the density of swarm. This research focuses on the problem of controlling the distribution of multi-agent systems considering a hierarchical control structure where the whole swarm coordination is achieved at the high-level and individual vehicle/agent control is managed at the low-level. High-level coordination algorithms uses macroscopic models that describes the collective behavior of the whole swarm and specify the agent motion commands, whose execution will lead to the desired swarm behavior. The low-level control laws execute the motion to follow these commands at the agent level. The main objective of this research is to develop high-level decision control policies and algorithms to achieve physically realizable commanding of the agents by imposing mission constraints on the distribution. We also make some connections with decentralized low-level motion control. This dissertation proposes a Markov chain based method to control the density distribution of the whole system where the implementation can be achieved in a decentralized manner with no communication between agents since establishing communication with large number of agents is highly challenging. The ultimate goal is to guide the overall density distribution of the system to a prescribed steady-state desired distribution while satisfying desired transition and safety constraints. Here, the desired distribution is determined based on the mission requirements, for example in the application of area search, the desired distribution should match closely with the probabilistic target locations. The proposed method is applicable for both systems with a single agent and systems with large number of agents due to the probabilistic nature, where the probability distribution of each agent's state evolves according to a finite-state and discrete-time Markov chain (MC). Hence, designing proper decision control policies requires numerically tractable solution methods for the synthesis of Markov chains. The synthesis problem has the form of a Linear Matrix Inequality Problem (LMI), with LMI formulation of the constraints. To this end, we propose convex necessary and sufficient conditions for safety constraints in Markov chains, which is a novel result in the Markov chain literature. In addition to LMI-based, offline, Markov matrix synthesis method, we also propose a QP-based, online, method to compute a time-varying Markov matrix based on the real-time density feedback. Both problems are convex optimization problems that can be solved in a reliable and tractable way, utilizing existing tools in the literature. A Low Earth Orbit (LEO) swarm simulations are presented to validate the effectiveness of the proposed algorithms. Another problem tackled as a part of this research is the generalization of the density control problem to autonomous mobile agents with two control modes: ON and OFF. Here, each mode consists of a (possibly overlapping) finite set of actions, that is, there exist a set of actions for the ON mode and another set for the OFF mode. We give formulation for a new Markov chain synthesis problem, with additional measurements for the state transitions, where a policy is designed to ensure desired safety and convergence properties for the underlying Markov chain.