Evolving and assembling functional link networks

Functional link networks (FLNs) are linear neural networks without hidden units whose ability to learn non-linear mappings depends on their being fed with suitable polynomial features. The discrete nature and huge dimension of the search space (subsets of polynomial features) clearly calls for an evolutionary approach. Our evolved FLN architectures (EFLNs) are derived by means of a genetic algorithm (GA) that imposes pressure on both classification performance and architectural simplicity. This gives rise to surprisingly simple and efficient networks such as those found for the Wisconsin breast cancer dataset. Further, it is shown that taking the majority vote of a reduced set of low degree EFLNs improves generalization significantly