A multiagent-based constructive approach for feedforward neural networks

In this paper, a new constructive approach for the automatic definition of feedforward neural networks (FNNs) is introduced. Such approach (named MASCoNN) is multiagent-oriented and, thus, can be regarded as a kind of hybrid (synergetic) system. MASCoNN centers upon the employment of a two-level hierarchy of agent-based elements for the progressive allocation of neuronal building blocks. By this means, an FNN can be considered as an architectural organization of reactive neural agents, orchestrated by deliberative coordination entities via synaptic interactions. MASCoNN was successfully applied to implement nonlinear dynamic systems identification devices and some comparative results, involving alternative proposals, are analyzed here.