On the configuration of multilayered feedforward networks by an evolutionary process

A learning algorithm based on genetic algorithms to configure multilayered feedforward networks in supervised learning mode is described. The method described lets a population of hidden units compete among themselves and "sell" themselves to be connected to members of another pool of output units. An output unit in this sense therefore represents a team comprising the output unit itself and its connected hidden units. This "team", of course, defines the architecture of the network. If each output unit can choose for itself how many hidden units it needs to accomplish the classification task, different architectures can be seen to be competing against each other. Experiment results are presented and discussed.