Well now I am at GMU and I am working on adding NEAT to ECJ! Back when I was learning about HyperNEAT I neglected to read about NEAT. Which now that I have been reading about it I wish I had taken the time. Then I would have known that what I called fault tolerance is called ablation testing and this testing was performed on NEAT!

I was accepted to present a poster about my research into the fault tolerance of HyperNEAT at the CCSC 2010 Conference at Juniata College. I put my research paper on my website at http://drew.freehostingcloud.com/research/hyperneat/fault-tolerance-of-hyperneat.html.

I created a capture the flag game to explore the capability of cooperative teams in a competitive environments using hyperNEAT.  I want to know if competing teams can learn when it is in their best interest to cooperate with a competing team.  I made a short video of my progress seen below:

Can an evolved hidden layer in the substrate handle faults not only in the output layer but also in the hidden layer without being trained with example faults?  I must first train it to get to the goal.  Then I should set one of the nodes in the hidden or output layers to be zero to test for fault tolerance.

If this doesn’t work I think that I need to move the output layer nodes to compensate since it is based on geometry.

I am currently experimenting with HyperNEAT to find out how well it responds to faults.  Such as, mechanical faults in a robot’s actuators.  I want to know whether HyperNEAT’s use of geometry will allow it to adapt to faults faster or slower than a regular feed forward backprop neural net.  I am using the CSharp HyperSharpNEAT implementation of HyperNEAT for my experiment.  By conducting this experiment I am, not only learning more about the algorithm, but I am also learning CSharp.

I just recently began to experiment with HyperNEAT (Hypercube based Neuroevolution of Augmenting Topologies). It is a new way to evolve neural networks created by Dr. Stanley. It uses the geometry of the placement of neurons in a neural network as input into a CPPN (Compositional Pattern Producing Network) which is evolved to learn the connection weights between neurons.