Co-evolving substrate layout in HyperNEAT

From Master Projects
Jump to: navigation, search

has title::Co-evolving the Artificial Neural Network substrate layout in HyperNEAT
status: finished
Master: project within::Computational Intelligence and Selforganisation
Student name: student name::Dimitar Nedev
Start start date:=2010/01/12
End end date:=2010/06/12
Supervisor: Evert Haasdijk
Second reader: has second reader::Gusz Eiben
Poster: has poster::Media:Media:Posternaam.pdf

Signature supervisor



One of the most complex, and certainly a very important, parameter of HyperNEAT is the topology of the substrate template for which the connection weights are determined by the HyperNEAT CPPN. We propose

  • to co-evolve a spatial CPPN that (similar to the picbreeder experiments) that yields a binary value for cells in a N-dimensional grid; for every value in the grid that receives a true value, a cell is placed. The issue here is to determine the extent of the empty grid, presenting us with yet another parameter for which to find a good setting.
  • to co-evolve a 'Bongard-like' controller for cell division, that allows to grow (and position - distance between cells means something in HyperNEAT).

In either case, the result is then a topology of cells, to be connected by the HyperNEAT CPPN. In both cases, we interleave evolution of the hyperNEAT CPPN and of the substrate-template generator. Questions to be answered: what interleaving rate should we set? Should both evolutions receive the same number of evaluations?