Evolutionary algorithm has recently shown a great possibility to determine the optimal structure of neural networks, but most approaches are based on the simple node to evolve. In order to facilitate more powerful neural networks, this paper proposes ...
Evolutionary algorithm has recently shown a great possibility to determine the optimal structure of neural networks, but most approaches are based on the simple node to evolve. In order to facilitate more powerful neural networks, this paper proposes another evolutionary method to develop neural networks based on modules, and shows the potential of the method by analyzing the behaviors of the modular neural networks obtained by evolution. The presented model might not only grow and evolve its own structure autonomously but also develop the cooperative functionality spontaneously. In the simulation of categorizing simple handwritten digits, we can observe that sophisticated network architectures as well as functional subsystems emerge from an initial set of randomly-connected networks. The performance itself is not comparable to practical pattern recognizers such as multilayer perceptrons and hidden Markov models, but the model can be thought of as combining connectionism and symbolic processing in terms of the functionality of each module.