Distributed Coordination and Estimation of Multi-agent Systems

Abstract

This thesis addresses problems arising in the coordination and estimation of a connected multi-agent system. The distributed coordination of agents consists of cooperative control and optimisation. Cooperative control aims at guiding a team of agents moving in a formation, maintaining a required pattern or velocity. An optimisation problem in a group of agents aims at driving them towards specific locations according to a performance index. Furthermore, the distributed estimation problem when states in a dynamical process are unknown is also considered; the agents are then assigned to estimate the unknown states cooperatively. A coordination problem investigated in this thesis is the coverage control problem in robot sensor networks whose objective is to find the optimal locations of the sensors leading to the best measurement. In this work, the objective function of locational optimisation is solved by interchanging the positions of the neighbouring agents. The formulated Lagrangian function augments the objective function and the consensus constraint to accommodate this mechanism. Accordingly, a coverage controller with cooperative constraint is proposed via the gradient-descent protocol. Furthermore, a controller is designed to drive the robots to the optimal locations in a finite time regardless of their initial positions. However, since an environment is not generally convex due to the presence of obstacles, a coverage controller with potential-field based obstacle avoidance is also considered to drive the robot to the optimal location safely. Motivated by the unavailability of information distribution in an environment before sensor deployment, a class of distributed nonlinear filter is designed to estimate the states of a dynamical process using the shared information among the agents. The proposed algorithm extends the distributed unscented Kalman filter to accommodate any communication topology. It utilises not only the measurement from an agent's sensor for estimating the process but also the shared information. The proposed filter is then employed to estimate the information distribution in the optimal coverage problem with unknown information distribution. Different from the existing field-estimation algorithms, the proposed filter also optimises the estimator gains in every iteration to avoid instability of the system caused by failure to choose the appropriate parameters. Thereafter, the question regarding how good is a distributed recursive Bayesian-based estimator when the information belongs to more general manifolds other than Euclidean, i.e., Riemannian manifolds, are also addressed. In this thesis, the formulated intrinsic Cramér-Rao bounds (ICRBs) demonstrates that the non-zero curvature term of the manifold also affects the performance of the Bayesian estimator. Lastly, assuming that the probability density functions (PDF) of the noises are Gaussian, a distributed nonlinear Kalman filter for Riemannian systems is also proposed. The simulations verify that the mean-squared error of the designed filter will not be lower than its ICRBs.

Date of Award	31 Dec 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhirun Hu (Supervisor) & Zhengtao Ding (Supervisor)