Figure 9: We pooled all neurons shown in Figure 8 and pooled all neurons shown in Figure 9, then applied CT clustering [16] to each pool in order to detect evolutionary trends over the different runs of the models. Clustering was performed on basis on the incoming connections only (as these are set by evolution only in both models). Each column shows one cluster of neurons. The pattern at the top of each column is an average over the downward connections of all neurons in the cluster. Below it are the upward connection patterns for each member neuron (for the main model, these are averages as shown in Figure 7(a)). Connection patterns were sign-flipped whenever this led to smaller distance to cluster center, to account for sign symmetry (see Figure 7(b)). (a) Result for baseline model. (b) Result for main model. The inset shows the triplet of neurons identified as the common solution evolved in 7 of the 10 runs of the main model. This triplet alone suffices for adequate behaviour.