Modular creation of neuronal networks for autonomous robot control

Abstract

In general, complex control tasks can be solved by dividing them into simpler ones which are easier to handle. Several authors have developed different solutions that combine Layer Evolution techniques with Evolving Neural Networks, giving rise to controllers made up by several networks. In this type of solution, the selection of the module to be used in each case is not an easy problem to solve. This paper is focused on a new evolutionary mechanism that allows combining modules which solve the different parts of a problem, giving place to a single recurrent neural network. In this way, simple modules which are trained independently of the problem to solve are used. The communication among them is established by evolution, which gives rise to a single neural network representing the expected solution. The proposed method in this paper has been used to solve the problem of obstacle evasion and target reaching using a Khepera II robot. The tests carried out, both in the simulated environment and over the real robot, have yielded really successful results.