Figure 3. A. Principle of the map-based, developmental encoding.

The neural network is encoded as a labeled graph (left), which is developped to a graph of maps according to the labels (right). (Methods). B. Principle of the HNN encoding (minimal HyperNEAT). Neurons are placed in a 3D substrate (top). To know whether two neurons are connected and the synaptic weight of each connection, a Compositional Pattern Producing Network (CPPN) is queried using the 3D coordinates of the two neurons (Methods). This CPPN is evolved using a direct encoding. To know the parameters of each node (neuron type and threshold value), a second CPPN is queried with the 3D coordinates of the neuron (Methods).