Evolutionary design of application tailored neural networks

An evolutionary algorithm for designing single hidden-layer feedforward neural networks is proposed. The algorithm constructs a problem-tailored neural network by incremental introduction of new hidden units. Each new hidden unit is added to the network by linear partitioning of the hidden-layer representation through a genetic search. A two-stage algorithm speed-up is achieved through: (1) a distributed genetic search for hidden-layer unit construction, along with the appropriate input to hidden-layer weights; and (2) a ´dynamic pocket algorithm´ for learning the hidden-to-output layer weights. Finally, promising experimental results are presented on the fast construction of small networks having good generalization properties.