I think there's a general disconnect between HN users and how the defense industry works.
When you are designing a weapons system designed to kill people, you cannot just design something 'in a hackish fashion'. No, this shit has got to work correctly all the time.
The cheapest Predator drone cost about $5M and top flight Predators and Global Hawks cost between $20M-50M. This means flying 100 of them is not reasonable. UAVs are also not autonomous and need manned pilots. Even if you were to build 100 drones, are you going to pay for 100 extra pilots, create some kind of airport that can house 100 drones, keeping them all fueled and maintained in the desert?
Kinda naive to think that you can just use SV start up philosophies in this application.
> When you are designing a weapons system designed to kill people, you cannot just design something 'in a hackish fashion'. No, this shit has got to work correctly all the time.
No, with that part of engineering as with EVERY part of engineering: it depends. Do you want UNIT reliability or SYSTEM reliability? Does every plane have a literal 100% chance of intercepting every target it is tasked with? No, of course not. But the missiles do, you say? Again wrong. But surely SOME portion of the entire system will work ALL THE TIME! Nope, no engineer worth his salt will give you a 100% guarantee for anything as unknown as "stopping all possible threats".
Furthermore plenty of systems that are supposed to be very reliable are made up of less reliable parts. The whole Star Wars program was based on defense in depth where no one layer of system was going to stop 100% of the warheads but multiple layers acting in concert would be able to (presumably) stop them all or almost all. Hackish is fine provided that you've got substantially overlapping coverage from multiple command and control drones. No one drone has to work 100% guaranteed because even at only a 95% success rate (which is abysmal compared to the "all the time" demand you're making) with three overlapping zones gives you .05 * .05 * .05 = .0125% chance of failing. I'll take 1/8 of a chance of failure per thousand incidents as successful enough.
Again, I'm not talking about having 100 drones with 100 pilots but perhaps having 100 drones with 9 pilots for the command and control drones (one for every 10 mules) and some kind of a very rudimentary, randomized loiter algorithm for the mules.
> Kinda naive to think that you can just use SV start up philosophies in this application.
I want to say a bunch of really snarky stuff in response. I'm an outsider to SV; I grew up in MN, went to school in FL and now I live in TX. None of these places ever get accused of being even SLIGHTLY SV-ish in nature so I don't see how I deserve that kind of comment. Ultimately though you've made a lot of assumptions that don't necessarily hold. The idea that you're going to defeat the US using the same kind of procurement and whatever that the US uses is a non-starter. You don't try and beat the US at a symmetric war but you can defeat them with an asymmetric one. We're losing how many lives and how much money in Iraq and Afghanistan right now to IEDs which are what, 10 notches below the fancy shit we have? And yet all our fancy airplanes haven't saved a single soldier from an IED that I'm aware of.
I'm not necessarily saying that you absolutely 100% can use SV philosophies in war and win, but I am saying that the notion that the way the US military does it is the ONLY way to do it isn't correct either.
I don't know why you are referring to the Star Wars program since it was not a program that was ever fielded.
Anyway, so is your strategy to fire multiple missiles at a time for each target? Assuming it's possible for UAVs to engage air targets (which they cannot right now), how many are you going to shoot off for each target, at the cost of $0.5M for each missile, just because you developed an algorithm in a 'hackish' manner?
If you gave each pilot 10 drones to control, good luck trying to execute evasive maneuvers on all of them when they come under attack.
Evasive maneuvers are something that a probabilistic algorithm tied into good sensors is probably better at performing for a swarm than a human would be, especially if the performance characteristics of these drones were made to be more extreme than most jets.
1. That's not a characteristic inherent to drones, that's just what we've done to date. And there are remote controlled fighter jets that the military uses for practice.
2. Because it seems well within our capabilities to do some basic missile evasion AI, that's just a small subset of the challenge of making a fully autonomous UAV. You need to identify the missile and its trajectory, identify its flight and kill characteristics, and then identify the various actions you could take and evaluate their likelihood of successfully evading the missile. Then you carry that flight plan out. Rinse and repeat until you're no longer in danger.
I'm not recommending that they do it one way or another - there're a lot of new possibilities that should be explored, which means that more of the same might not be the way to go. All I was saying that it's silly to say categorically that "drones are slow", because they certainly don't have to be - if anything, they have the potential to be even faster and more maneuverable than manned fighter jets, because you can get rid of a lot of the cruft and restrictions related to keeping a pilot alive and letting them pilot effectively. No more canopy, no more cockpit shielding, no more visible instrumentation and layout requirements, no more ejection seat, no more oxygen supply, no more human survivable g-limits, etc.
I'm also not advocating for a completely autonomous drone, just saying that automatically flying a plane to evade a missile doesn't seem terribly difficult, as far as AI problems go, and the amounts of money thrown at these sorts of things are absurdly large.
Heh probably at least a little. I've got a decent grounding in AI and some in computer vision, and those portions at least seem doable, I've worked on some basic guidance algorithms before, and I know people who are having some success making generalized autopilot systems for drones for spatial waypoint following and flight controls, so that portion seems achievable as well. Once you can reach that level of abstraction, it starts looking like something a minimax algo with sufficient look-ahead depth could deal with, at least with a very rough level of precision. But yeah, devil and details.
One issue you're not considering WRT to the Strategic Defense Initiative (SDI, vs. "Star Wars" as used by its opponents) is that even a partially effective system negates an adversary's first strike capability, since they can't choose which missiles and warheads get taken out.
When you are designing a weapons system designed to kill people, you cannot just design something 'in a hackish fashion'. No, this shit has got to work correctly all the time.
The cheapest Predator drone cost about $5M and top flight Predators and Global Hawks cost between $20M-50M. This means flying 100 of them is not reasonable. UAVs are also not autonomous and need manned pilots. Even if you were to build 100 drones, are you going to pay for 100 extra pilots, create some kind of airport that can house 100 drones, keeping them all fueled and maintained in the desert?
Kinda naive to think that you can just use SV start up philosophies in this application.