Neural Network MIssile Guidance

Would it be feasible to train a neural network of some kind that is inside of a missile, to recognise the silhouette of a plane in the sky, and manouevre itself so as to colldide / detonate within a kill-radius of the target? If possible, this would mean to training for operator, FIre-and-forget, immune to most if not all countermeasures, independent of weather and unjammable.

>>14602 No, not really. You could theoretically disable a missile of that type by firing a laser at it and blinding it

there's already a paper on that and I would assume there's a patent

>>14603 continurally shining a laser at a tiny lense that's travelling at mach 4. mmmm also there'd be no warning, unlike other types where the target is continaully designated i.e laser guidance. thus deploying a countermeasure of any sort would be even harder.

>>14605 Hard, but not impossible

>>14602 blocks your path. A RPI can already do OCR with some additional chips. So that would be absolutely no problem. To be honest That does not sound like a major improvement to current tech. It will be sensible to anything visual. Flares and making a smoke go of like in parades will also jam it. What currently strikes me as an interesting way to use more powerful computers would be to try to identify flares as flares. To make them no longer the new sexy heat source to blow up. Flares do not burn jet fuel, or even if not in the same conditions. Even if they glow similarly it is probable that a trained NN would be able to differentiate quickly. (recognizing texture or emission spectrum). You could probably record data on parades (for the plane signatures) and a few weeks days of first attempts to shoot enemy planes down.(for the flares signature) Then you will have quite a chance to turn a flare thrower a pile of ashes.

>>14602 >unjammable >immune to countermeasures Eh. Neural nets aren't magic. In fact, adversarial attacks are very easy. You can put a one-inch sticker that looks like a stop sign on something on the side of the road and it'll freak out self-driving car neural nets. And they run off standard cameras. Blind it with chaff or a laser or any kind of even medium-intensity light and it's worthless. A missile tracking down a plane in an empty sky is kind of a simple task for classical algorithms. Hardly any complexity; no need for neural nets. Especially when the plane is absolutely lit up on infrared. Neural nets are, for the most part, black boxes and are thus hard to debug or weed out unintended behavior based on unknown stimulus. This is changing, but it's still a pain.

>>14607 >>14610 Interesting. Yeah adding a modicum of intelligence to infrared-homing missiles to reduce likelihood of being tricked by flares sounds doable and worthwhile. No reason to not have the missile "see" in IR, and let its little brain figure out the rest. For some reason I thought the thermographic sensing part was clunky and expensive hell compared with a "normal" camera, and would be nice to have bypassed.

>>14610 >And they run off standard cameras. Blind it with chaff or a laser or any kind of even medium-intensity light and it's worthless. "Yes, it gets jammed easily, I'm afraid. It's all made from currently-existing technology."

Pffff. What is this soykaf? Transistors? Electricity? All we need is a firing computer and a few good M E N . https://m.youtube.com/watch?v=gwf5mAlI7Ug

>Pffff. What is this soykaf? Transistors? Electricity? All we need is a firing computer and a few good M E N . And in 50 years? curmudgeons like you'll be saying "All we need is a neural-net and a few good M E N .". There's no such thing as humor

>>14614 That's a good point, yeah that could work. Probably already have systems like that somewhere in the world. And yes, IR cameras are very cheap if you don't care about the quality or band of IR. Basically you just take an inexpensive webcam sensor and remove the filter or lens coating which cuts out near IR. That's how cheapo night vision security cameras work--the kind that require a ring of IR LEDs to illuminate their field of view. But something belching lots of wide-band IR like a jet engine would show up just fine on it. Wouldn't really need something as sophisticated as FLIR or an IR bolometer. >But I bet the US military uses the expensive ones in missiles anyway, because Raytheon >>14620 Somewhat related, but they did briefly experiment with missiles which were guided by pigeons. They trained the poor birds to sit in a special cradle in the nose cone and peck at a sensor in the direction of a target for a food pellet reward. The reward was of course short-lived. https://www.youtube.com/watch?v=Ro068H6w8Vg That might count as a missile with a neural net?

>>14620 Pffff. This is a high tech board for high tech ideas

>>14639 wow. that pigeon guided missile is ingenious.

>>14644 Right? Before microcontrollers people still managed to find a way. Pretty astounding when viewed from a time when the solution to anything involves several layers of software. I think it's useful to remember sometimes.

would it be feasible? Sure would it be practical? Hell no. There are way better guidance systems that don't rely on a slow unreliable neural net. And in missile systems, speed and reliability are paramount.

>>14667 I think those types that you consider "better" probably leverage a lot of operator training, and would therefore necessarily not be fire-and-forget. It kind of goes without saying that any missile guidance that expects a highly trained soldier to continuously aim a laser at the target --or literally control the missile with a controller in his hands vis radio and even physical wire-- wouldn't have much to gain from AI, as the intelligence in those circumstances is provided by a human; expensively trained and still rather liable to miss the target. i suppose talking strictly only about fire-and-forget missiles, I think the only option for improving existing technology would be incorporating some kind of ML.

>>14678 Why do you think that the only way to implement a fire-and-forget missile would involve ML? This is the problem with computer science today. People expect ML to be a silver bullet and fix everything, then they think it's the only way to do anything, and they forget about far more practical, reliable, and powerful traditional algorithms. ML has serious problems once it leaves the lab. Also not sure why you think a non-ML missile would require a highly-trained operator. There's a lot of other options besides haphazardly-trained AI and a radio-controlled missile that a person has to literally control with a joystick or a laser pointer. You can get a very reliable IR/laser/visual image/radar lock and maintain it and steer the missile's control surfaces without any kind of ML or neural net.

>>14679 when you say "maintain it", what is doing the maintaining? an operator on the ground constantly re-designating the target?

>>14604 Source please. No. I do not think NNs/ML is adequate for a high-performance environment like missile guidance system. One needs to have both provably and verifiably correct system with real time constraints [1]. ML and NNs tend to still mess up, and as >>14610 states, can easily be duped. Going back to the real time system constraints, I am not sure if any of you here has actually worked in a real time system (for example drones), but we really need to be sure of everything to the point where most real time environments (such as Ravenscar Profile [2] for Ada) tend to remove a lot of features for the sake of meeting time constraints. As an example, short-circuiting boolean expressions is not great for real time systems as you can easily add uncertainty to your system. NNs and MLs do not follow the above. We can not prove and ML algorithm can and will identify the target in the time as maybe the angle of attack doesn't allow it to register the target correctly. Plus in a highly-performant environment like a missile, one absolutely cannot compromise on speed [1]. Why do you think most embedded run-time environments are so strict and featureless in comparison to their general counterparts [2]? With respect to optical dazzlers, yes current implementations of missile guidance techniques often fall for them, but I definitely don't think the solution is ML algorithms. Few of the reasons stated above. >>14678 You have not expanded on why you think that. >>14686 Not OP but those are Television-guided missiles [3]. Not that relevant to fire-and-forget ones. References: [1]: https://en.wikipedia.org/wiki/Real-time_computing [2]: https://en.wikipedia.org/wiki/Ravenscar_profile [3]: https://en.wikipedia.org/wiki/Television_guidance

>>14689 Same dude here. Just forgot to mention that NNs and ML algorithms also tend to have (sometimes unforeseen) edge cases which can result in catastrophic failure in the case of a missile. Also wanted to correct a typo. > We can not prove an ML algorithm can and will identify the target in the time constraints

>>14686 A microcontroller in the missile, of course. An operator on the ground?? What year is it? You can have autonomous computer control without ML, you know. I think this might be getting closer to the root of this misconception. A regular old program can be written to keep an object in the sensor window, especially when the object is surrounded by empty space. There is no need for neural nets in such a simple application.

https://en.wikipedia.org/wiki/PID_controller Might want to read up on this, OP. What you want for a missile is a very fast real-time PID loop with provable responses and provable failure modes. The fast and provable parts are where NN and ML fail pretty spectacularly, at least as of 2019. People are working on it, but expect reasonable incremental progress--not miracles. For example, you could have a very stripped-down and optimized traditional algorithm that rapidly scans the sensor window and calculates the center of intensity of the target, then calculates the position of that center of intensity relative to the edges of the sensor field. Then the program would implement a change in the control surfaces of the missile using PID to adjust attitude without overshooting or introducing any harmonic wobble and bring the target closer to the center of the sensor window. Wash and repeat as many times a second as your microcontroller is capable of doing accurately, and you'll have a very difficult to evade self-guided missile. If you wanted a missile that's even harder to shake, you could introduce some galvo-based optics to make the sensor able to keep track of a squirrely target far more responsively than the missile itself is capable of maneuvering, but this is getting into details. There are also ways to make a traditional program tend to reject adversarial input like chaff--that would require getting more creative and would necessarily slow down the control loop a lot, but it's still going to be way faster than a neural net. All of this would be regular ol' code that can be read by humans, debugged, tested with automation, etc. in ways that an ML black box never could. At least not yet, anyway. And it will be blindingly fast if done properly, and will respond in a predictable way to any stimulus. Whereas you can never actually prove a neural net won't have a certain input that causes it to utterly freak out. There's no known way to exhaustively search a neural net's response to unexpected stimulus. In this way, it would most likely be easier to distract with chaff and ECM than a traditional algorithm instead of being more robust. Especially once enemies start optimizing for defense against these hypothetical ML munitions. That might produce some interesting tactics, actually. As another poster said, drone design and programming involves a lot of the same fast control loops that you'd expect to find in basically any other avionics control. Avionics is a very picky field. I mean if you exclude Boeing's recent moronic bungling, but that should serve as the exception to prove the rule. Especially military avionics; they are going to want that code to be clean as a whistle, and provably so. They will want to know to several decimal places of certainty that the algorithm isn't going to bodge up, fly away and lock onto a friendly or a civilian airliner or a shopping mall on the ground or whatever else. You can't really do that with ML at this juncture.

>>14693 I don't know about missiles, but I currently program torpedoes at work, and these are way "smarter" these days, since they must face a lot of countermeasures (jammers, decoys) and a not so simple environments near coastlines. There's some ML (Bayesian) stuff in the target classification process.

>>14695 That's interesting. Where does the Bayesian ML come up in the process? Or I suppose you might not be allowed to talk about stuff like that, don't van yourself or anything