A similar complex, but less studied problem is the size and composition of the Training Data Set (TDS) for (sub)optimal ANN training. We present an Evolutionary Resample and Combine (erc) selecting TDS patterns in parallel out of all available data by optimizing the generalization capabilities of a fixed architecture ANN.

Quite naturally the above approaches can be combined to coevolve ANN architecture and TDSs. We employ symbiotic coevolution where independent populations of ANNs and TDSs form each others environment. As ANN fitness is equally credited to the TDS it has been trained with, the emergence of robust structures can be observed. A simple pattern recognition problem, benchmark problems, and classification of satellite imagery are used to evaluate and compare the various methods.