ALENA : Adaptive-Length Evolving Neural Arrays

Evolving neural arrays (ENA) have proved to be capable of learning complex behaviors, i.e., problems whose solution requires strategy learning. For this reason, they present many applications in various areas such as robotics and process control. Unlike conventional methods "based on a single neural network" ENAs are made up of a set of networks organized as an array. Each of them represents a part of the expected solution. This work describes a new method, ALENA, that enhances the solutions obtained by solving the main deficiencies of ENA since it eases the obtaining of specialized components, does not require the explicit decomposition of the problem into subtasks, and is capable of automatically adjusting the arrays length for each particular use. The measurements of the proposed method "applied to problems of obstacle evasion and objects collection" show the superiority of ALENA in relation to the traditional methods that deal with populations of neural networks. SANE has been used in particular as a comparative referent due to its high performance. Eventually, conclusions and some future lines of work are presented.
Facultad de Informática