The determination of permeability is an example of many geological pro
blems where laboratory-measured data is expensive and limited in quant
ity. We related permeability values to well logs. We used neural netwo
rks trained both with the popular backpropagation algorithm and with a
genetic algorithm. The genetic training produced smaller errors and b
etter generalization than backpropagation training on the same network
topology. The cost includes,greater average computation time as well
as,greater variation in computation time for the genetic training. The
genetic training is robust and not sensitive to selection of the cros
sover and mutation parameters.