wouldn't back propagation (when time is better) be better?

Hi again,

I'll try to answer your question, but I hope a NNs guru will pop in and give you a better answer :)

I think backprop requires you to know what the output should be. In this case, since you're talking about the whole lap time, you don't have a specific output to train the network but a whole array of outputs that gets executed over the lap time.

What I'd recommend then, is either a), use the GA solution with the lap time, or b), write a few sets of inputs with their matching outputs, and train the network with that using backprop.

Also, if you're interested in GA/NNs, check out NEAT.

Hope this helps

Eric

edit: ack, typos