Distributed decision-making of networked multi-agent systems in complex environments

This dissertation is concerned with distributed decision making in networked multi-agent systems; that is, developing practical mechanisms which agents can utilize to autonomously coordinate their actions/decisions through local message exchanges and successfully achieve a system level goal with a satisfactory performance guarantee. In particular, this dissertation is divided into three parts and each one focuses on the following three classes of problems : (1) distributed average consensus; (2)distributed cooperative constrained optimization; (3) distributed online learning based coordination. This dissertation starts from the fundamental problem of distributed average consensus. In Part I, we first propose a class of dynamic average consensus algorithms and show that these algorithms allow agents to asymptotically track the average of a class of time-varying individual reference inputs. We then come up with a class of gossip- based algorithms which agents can use to achieve approximate average consensus via exchanging quantized information. Part II is concerned with a class of general multi-agent optimization problems. In particular, each agent is associated with a local objective function and a local constrained set. There is a pair of inequality and equality constraints known to all the agents. We first present a class of distributed primal-dual subgradient algorithms to solve the case when all the ingredients are convex. We then introduce a distributed approximate dual subgradient algorithm to address the non-convex counterpart. Part III studies distributed coordination schemes with online learning. The first problem considered is to optimally deploy a group of visual mobile sensors where the environmental distribution is unknown a priori. We formulate the problem as a non-cooperative game and come with up two distributed learning algorithms which allow sensors converge to the set of Nash equilibria and global optimum with probability one, respectively. The second problem is distributed formation control against a class of deception attacks. We propose a class of algorithms which allow vehicles adapt their strategies online and achieve the desired formation in the presence of deception attack