Controle descentralizado para formação de sistemas multi-agente baseado em programação dinâmica adaptativa e polígonos regulares

Abstract

In multi-agent systems, control is decentralized when decision-making is done by individual agents and not by a centralized unit that processes the states of all agents. The main difference in information/data processing when working with centralized control versus decentralized control, more precisely when solving the

dynamic system, is the absence of the Laplacian matrix that presents the char- acteristics of interconnection between the agents of the system, with its absence,

the system does not have complete information about the number of agents and

the connection mode between them. In order to overcome this lack of informa- tion, it is proposed to use shapes that have well-defined characteristics (sides and

angles) such as Regular Polygons. Due to the use of these geometric figures, the

shape of the multi-agent system is defined by the number of agents and the de- sired distance between them. In this methodology, there is no need to exchange

information between the agents, since each agent works only with the distances of its neighbors for the assembly of the shape, which are easily decomposed to

the Cartesian plane. To ensure the optimality and adaptability of the control sys- tem, the controller gain update is obtained by heuristic dynamic programming

approaches, which is executed in parallel to the main algorithm proposed in this thesis. This methodology is evaluated for a formation of three and four mobile agents.