CAP6938
Neuroevolution and
Artificial Embryogeny
Approaches to
Neuroevolution
Dr. Kenneth Stanley
February 1, 2006
Many TWEANN Problems
• Competing conventions problem
– Topology matching problem
• Initial population topology randomization
– Defective starter genomes
– Unnecessarily high-dimensional search space
• Loss of innovative structures
– More complex can’t compete in the short run
– Need to protect innovation
• How do researchers design NE methods?
Breeder Genetic Programming
(Zhang and Muhlenbein)
• Represent network as a tree (TWEANN)
• Only crossover adapts topology
• Attempt to minimize both complexity and
error:
• Tested with parity and majority functions
Parallel Distributed Genetic
Programming (PDGP)
Pujol and Poli (1997)
• “Dual representation”: linear and graph
Parallel Distributed Genetic
Programming (PDGP)
Pujol and Poli (1997)
• 2D genome uses
overrepresentation
• Several crossover operators use
properties of both 1D and 2D
representations (e.g. subgraph
swapping)
• Also several mutation operators
• Fixed-sized genome
• Also tested on parity (and later
control)
GeNeralized Acquisition of Recurrent Links
(GNARL)
Angeline, Saunders, and Pollack (1993)
• “Thus, the prospect of evolving connectionist
networks with crossover appears limited in general,
and better results should be expected with
reproduction heuristics that respect the uniqueness of
the distributed representations.”
• Random initial networks
• Fixed-sized genomes
• Structural mutations
• Tested with “Inducing Languages” and “Ant Problem”
Structured Genetic Algorithm (sGA)
Dasgupta and McGregor (1992)
• “Standard” matrix representation
•
•
•
•
Size of matrix is square of # nodes
Maximum net size for fixed matrix size
No thought to crossover (just plain GA)
Tested on “multi-solution functions”
Cellular Encoding
Gruau (1993, 1996)
• Indirect encoding (Artificial Embryogeny)
• First method to balance 2 poles without
velocity inputs
• Biological motivation: grow from single cell
• Gruau proved CE can generate any graph
• Crossover swaps subtrees like GP
• Indirect encoding only makes competing
conventions harder to comprehend
Cellular Encoding
Gruau (1993,1996)
Enforced SubPopulations (ESP)
Gomez and Miikkulainen
(1997,1999)
• Fixed-topology successor to Symbiotic Adaptive
NeuroEvolution (SANE; Moriarty and
Miikkulainen 1996)
•
•
•
•
Neurons evolved in subpopulations
One subpopulation for each hidden neuron
Cooperative coevolution
Interesting circumvention of competing
conventions
ESP defeats CE
Hidden Nodes
Inputs
(Gomez and Miikkulainen 1999)
TWEANNS need Principles
• Is there a principled method for evolving
topologies that is not ad hoc?
• How can the TWEANN challenges be
handled directly?
• Are all TWEANNs created equal?
•
•
•
•
Next Class:
NeuroEvolution of Augmenting
Topologies (NEAT)
Directly address TWEANN challenges
Turns topology into an advantage
Applicable outside NN’s
Basis of class projects
Evolving Neural Networks Through Augmenting Topologies by Kenneth O. Stanley and
Risto Miikkulainen (2002)
Homework due 2/6/05: 1 page project proposal including project
description and goals, a falsifiable hypothesis on what you expect to
happen, why it involves structure, and what platform you will use
(language and OS). If partners, describe briefly division of labor.