Lethal Autonomy: Machine learning, operation and predictability. Can we truly control it?

How can the behavior of autonomous weapons be trained, guided and ultimately constrained? Predictability and controllability are key when it comes to the battlefield, but is it given without human oversight?