Neural Nets

I've been researching neural nets for several years and while I feel that I have a decent grasp on the subject matter, I have not yet produces an interesting demo.

Connect 4

Todo:

• Select only a few of the top champions for mutation - there's not point to creating a mutation of 1 when we're up to 1000
• Break the make table function off to it's own process
• Possibly create a more competitive model that NNs can reproduce when they win and will die after so many losses

Genetic research

I need to experiment with genetic evolution. Instead of evolving the neural networks directly, a string that encodes how to build a neural net is evolved. This string includes genetic markers so that individual structures can be evolved and tuned and those structures can be combined.

Stack

• Perceptron evolution: Only input and output neurons, evolve the optimal weights here.
• Genetic evolution: Individual sets of genes are evolved
• Genetic combination: Genes are combined and evolved
• Sexual reproduction: Sets of genes are compared and combined
• Tuning: In the final phase, the raw values of the neural net are evolved. I believe coming into this level, there will be a lot of zeros in the grid, tweaking those values a little could produce some fine tuning.

File:Floreano-neuroevolution-overview-2008.pdf File:CheckersANN.pdf

Also see

• Designing neural networks by genetic algorithms using graph generation system