spoffojj@hq.af.mil (Jason Spofford) (08/01/90)
I have been conducting research into evolving neural networks with a genetic algorithm (GA). For now I have been evolving neural networks that solve straight forward pattern recognition problems. I am fairly new to the neural network world and I seek a little feedback on my initial results. I've written a 4,000 line C program that implements a GA. I use a simple neural network model with hard-limiting output and integer weights. I've developed a genetic code that allows weights and thresholds to be modified as well as the neural network connection structure. Feedback and feedforward strutures are allowed. One limitation is that neurons can only have four inputs, no more, no less. Let me get on to the results. In one experiment, my program generated neural networks to recognize all the letters in the alphabet (capital letters only). Each network had 25 inputs (in a 5x5 binary array). My program (Evolve), generated a ten neuron network, with feedback, that recognized all 26 patterns correctly. BTW, in the GA, smaller networks were favored over larger ones of the same performace. Does anyone know how many neurons it would take for back-prop to do the same? Another experiment, larger in scope, involved a 9x9 binary array of 37 patterns. Again, Evolve developed a network to recognize all the patterns. The genetic code was on the order of 6400 bits long! The solution networks had 91 neurons. I appreciate any comments regarding these results. -- ---------------------------------------------------------- ) Jason Spofford <((((((> spoffojj.hq.af.mil ( ) ( ----------------------------------------------------------