Comments by "TheEvertw" (@TheEvertw) on "Sidewinder - The Weapon That Changed Air Combat" video.

keith moore Have you noticed the maneuvers made by the last sidewinder in the video? That dance involved an awful lot of decision making! Alternatively, as all robotic behavior has been somehow programmed or trained into it, one can argue that no robot has the ability to decide anything.
10
keith moore "a robot would be able to tell" Only if you, the programmer / trainer, told it very clearly how, and what to do in either case. For a robot, the task to differentiate between a cargo or fighter plane is little different from optimizing its flight path to intersect with its target.
8
The most important tool of the engineer is her/his brain.
4
@saudade2100 "They went to the moon" And back again! That was the hard part. (In Dutch, "they / it went to the moon" means the same as "it all went to pieces")
4
@letsburn00 On the other hand, modern engineers don't need to spend days upon days tediously working through equations or hand-crafting prototypes. Computers allow them to investigate the design space much quicker. And I mean MUCH quicker. The modern focus on core work has more to do with the way projects are managed nowadays than with anything else.
3
@rodanone4895 "have you studied fuzzy/neural systems" I have indeed. fuzzy controllers are fine in this context, but I have problems with neural networks controlling weapon systems due to their inherent non-determinism. First there needs to be a legal framework where a neural network (and the company that created it) that caused collateral damage is treated equivalently as a human that does. Humans make mistakes, and so do neural networks. Re: modern Aim9's: Likely these do not use AI / neural networks but instead use some form of predictive control, probably a robust variant. They probably calculate the likelihood their target will be at a certain point in space as a certain time based on observed target behavior, and optimize the probability of intercept, perhaps even using a model of expected evasive maneuvers & counter measures. At least that is how I would do it. As this involves a lot of classification of observed behavior, there might be some opportunity for neural networks to detect e.g. the ejection of Flares and picking the target out of these flares. But I would not let them control the overall flight path planning. Using neural networks as black boxes that get all the sensor data and control all actuators is far too uncertain in my opinion. But to use one or more networks to solve sub-problems of limited scope with well-defined situations seems promising. By limiting the scope of the problem a net has to handle, they become orders of magnitude easier to train.
1