There are 4 hidden nodes used. The process goes through 20 rounds with 20 neural network populations. If necessary, the next 20 rounds will be processed with a doubled number of hidden nodes, which is 8. However, in this example there are no rounds with 4 hidden nodes failing, so using 4 hidden nodes is enough.
The population with the smallest checking error is selected. It evolved over 32 generations. At each generation, 10% of neural networks are selected according to the mutation rate, which are learned in 100 cycles in the direction of reducing the squared mean error on the training samples. In our example, the best neural network of the best population is chosen as the destination neural network. It has the final learning coefficient of 0.246852. Other activities of the populations are hybridization and mutation of weights according to the general method of genetic algorithm.