The first part of this work deals with the problem of positioning a swarm of relay agents with the objective of providing connectivity to a second group of active agents (pursuers). The first approach consists of modelling the considered scenario as a quadratic programming problem (QP) with linear restrictions, using a fixed graph structure. In order to solve such model, a recurrent neural network is proposed with fast convergence rate to the optimal solution, even with reasonably big size instances. In addition, a comparison with the Matlab QP solver has been conducted in some experimental simulations. In the second approach, it is proposed an estimation of the algebraic connectivity of the underlying graph generated by the network. Over this estimation, it is proposed a metric to direct the group of relays and pursuers, using only local neighbourhood information of each agent. On this approach, the graph structure is dynamic and it is also proposed the use of the schedule by edge reversal (SER) as a solution to ordering the parallelization of the robot positioning computation. Moreover, the meta-heuristics Simulated Annealing, Genetic Algorithms, Particle Swarm Optimization and Backtracking Search Algorithm have been applied as alternative solution providers. The proposed solutions have been applied in a target pursuit scenario, for which the targets are deployed in different spots and may have some reactive behaviours (such as escape from the pursuers).The second part of this work investigates the visual tracking of shape shift objects in real time. As a solution, it is proposed a short- and long-time memories tracker which uses a weightless neural network for training and retraining the objects patterns.
|
