I have been very interested in genetic algorithms for some time now and have enjoyed playing with them and trying to improve them. My most recent project involves using a leading genetic algorithm package (NEAT, from UT Austin) and trying to automate optimization for the large number of parameters required for the program. I did this by using the genetic algorithm on itself to find an optimal configuration of parameters to match randomly generated networks. By altering the type of networks generated, I was able to observe the optimal setting for each parameter under a variety of conditions.

Before using NEAT, my attention focused on sensor nets and how they could be improved using genetic algorithms. I constructed 3D simulations of scenes with people and sensors in them using DirectX. From these scenes I attempted to train a network to determine the location of each person. Although I didn't succeed, I gained valuable experience working with DirectX and insights into the weaknesses of genetic algorithms.

One of my earliest experiments with genetic algorithms was an attempt to see how easily a network with the ability to do one-shot-learning could evolve. The experiment evolved simple networks of idealized neurons. These networks were trained on input patterns and given fitness based on how well they could recall the original input given only a partial pattern. As part of the experiment I used simulated annealing to regulate how quickly the genome would change. The networks that evolved didn't meet my expectations, but I would like to believe that this was due to a lack of computational power (limiting networks to only a few dozen neurons) rather than a flaw in my framework.