I was looking through the tutorials here, in particular using a NN to control the yaw of a bot to avoid obstacles. I was wondering if the same could be applied to controlling a ship via keypresses. For instance assume the ship is locked in a 2d plane (of a 3d space in my case) and the ship can go forward, backward, turn right and turn left. Forward/backward is speed increase and decrease. Would 4 NNs, one for each button, actually work to control the ship? Given the tutorial it seems like they might if they all trained together. And could this be used to fire weapons as well (I know this is a broad question but you get the idea, I hope :). So I thought maybe training the NNs under various circumstances based on the possible high level states of the agent would be one way to use them. For instance you train 'attack' nets for an aggressive response and maybe 'flee' nets that causes the agent hide behind obstacles and try to get away. Though I'm not quite sure if the different nets would be able to coordinate actions correctly. Though while I can see it working for movement I'm not sure it would work for firing weapons in my case since weapon placement is effectively random (not to mention there are firing angles that constrain when weapons can fire), though I guess NNs for each type of ship is a possibility :)
Thanks,
Eudaemon

I'm confused as to why you want the human to control an AI layer, then the ship. Couldn't you just map the key-presses onto movement?

If you use different neural networks, try evolving them separately -- you'll get better results & performance. I'd stick to 1 NN per behaviour, rather than multiple small ones...

What's the plan?