Comparison of NEAT and HyperNEAT on a Strategic Decision-Making Problem

Created: 2022-01-28T20:32:29-06:00

Return to the Index

This card pertains to a resource available on the internet.

• Neuron genes assign an ID and neuron type in the network.
• Link genes associate neurons to one another (dendrites.)
• Dendrites have an "innovation number" for the order the links were inserted. If a link is removed and later re-added it is reassigned its original innovation number.

Structural mutation

• Can mutate by adding or removing some randomly sampled amount of weight from a connection.
• Can connect two nodes that were not connected before.
• Can split a node by inserting a neuron between an existing dendrite.

Speciation

Speciation divides candidates in to families who are close to their original progenitors. Fitness is shared by family to allow changes the chance to survive their initial lack of tuning.

• Entities which are sufficiently similar to the original member of a species are placed in a bucket with all other sufficiently similar members.
• Entities too different become the originator of a new species.
• Innovation numbers are used to track the origins of species.
• Fitness is shared across a species.

HyperNEAT

Compositional Pattern-Producing Networks

Creates a hypercube where evaluation at a given space determines the existence and intensity of connections

Samples the hypercube to produce the neural network

Along other lines, it has been demonstrated that machine learning techniques such as random decision forests can be made to determine for themselves when they have had enough training through the use of statistical hypothesis tests, allowing them to learn more quickly using sparse data (McBride et al, 2009).