Max Wolf's Second Brain

Evolving Neural Networks That Are Both Modular and Regular - HyperNeat Plus the Connection Cost Technique

1 min read

year: 2014/07
paper: https://www.researchgate.net/publication/262013945_Evolving_Neural_Networks_That_Are_Both_Modular_and_Regular_HyperNeat_Plus_the_Connection_Cost_Technique
website: Presentation | short 5min high level vid
code:
connections: HyperNEAT, modularity, regularity, connection cost, Jeff Clune

Connection cost is added as a third optimization objective (alongside performance and behavioral diversity).
The connection cost they use is the sum of squared lengths of all connections in the phenotype network.

The paper tests four main algos:

  1. HyperNEAT - Standard HyperNEAT with Link-Expression Output (LEO)
  2. HyperNEAT-GS - HyperNEAT with Gaussian Seed (includes LEO + a Gaussian activation node to encourage local connections)
  3. HyperNEAT-CCT - HyperNEAT with Connection Cost Technique (includes LEO + connection cost as optimization objective)
  4. DirectEncoding-CCT - Direct encoding with Connection Cost Technique (comparison baseline from previous work)

Graph View