Hands-On Neuroevolution with Python

Compositional Pattern Producing Networks

HyperNEAT extends the original NEAT algorithm by introducing a new type of indirect genome encoding scheme called CPPNs. This type of encoding makes it possible to represent the connectivity patterns of a phenotype's ANN as a function of its geometry.

HyperNEAT stores the connectivity pattern of the phenotype neural network as a four-dimensional hypercube, where each point encodes the connection between two nodes (that is, the coordinates of the source and target neurons) and the connective CPPN paints various patterns within it. In other words, CPPN computes the four-dimensional function, which is defined as follows:

Here, the source node is at (x₁, y₁) and the target node is at (x₂, y₂). At this ...