Multi-agent State Estimation

The knowledge about its operation environment is a fundamental requirements for intelligents agents acting in a dynamic world. Knowledge gathering is thus a critical functionality for any agents and it involves several problems: perception, object detection, environment structure reconstruction and so on. The state estimation problem is a general definition to describe many applications: object tracking, localization, mapping, exploration. The core concept is that the agent has to operate in an environment, basing its actions on what it thinks about the environment. The “state estimation problem” has this goal: given some kind of sensations, it reconstructs a model of the world. Sometimes an agent can have an idea “a priori” about how its world is made. But often it does not know anything and has to also build a model on data. While a lot of researches has been done in the past about single-agent state estimation problem, in recent years, a lot of reserchers focused their attention about exploiting state estimation technique using multiple agents. Using several agents gives more challenging questions. For example, should agents share their own informations or not? How deal with communication problems? And what about scalability of methods? All of the state-of-the-art approches virtually use Bayes filter. This choice is necessary because using agents based on unreliable sensors, that gives uncertain measurements, we could easily have unbounded error solutions. This paper is structured as follow. In Section 2.1 we first describe a general bayesian framework, and we describe two of most used techniques: the Kalman filter and the particle filter. Next, in Section, 3 we present an overview of state-of-the-art approaches, from centralized methods, that use a single node to gathering information, to distributed platforms, that use instead the capabilities of distributed computation. Finally in Section 4 we point out some interesting results in current devoloping architecture.