Non-redundant genetic coding of neural networks

Feedforward neural networks have a number of functionally equivalent symmetries that make them difficult to optimise with genetic recombination operators. Although this problem has received considerable attention in the past, the proposed solutions all have a heuristic nature. We discuss a neural network genotype representation that completely eliminates the functional redundancies by transforming each neural network into its canonical form. This transformation is computationally extremely simple, since it only requires flipping the sign of some of the weights, followed by sorting the hidden neurons according to their bias. We have compared the redundant and non-redundant representations on the basis of their crossover correlation coefficient. As expected, the redundancy elimination results in a much higher crossover correlation coefficient, which shows that more information is now transmitted from the parents to the children. Finally, experimental results are given for the two-spirals classification problem.