[lucasjung]

Nothing you say addresses my broader point that there are currently too many situations in aviation where the complexity of the decisions involved exceeds our current capacity for automation.

A billion dollars was spent by the eu in the 80s on self-driving cars. They didn't completely succeed. It looked like noone would succeed for 30 years. And yet, bang, when the competition's opened up, a couple of guys from stanford do it. I believe the technology to solve the problem is out there, it's just a matter of the right people trying at it.

If they tried in the '80s and the guys from Stanford did it in the 2000's, then it was almost 30 years before anyone succeeded. I think that success had a lot more to do with technology maturing over time than it did with "the right people trying at it."

>Computer vision is close to being solved. Look at kinect, kinect 2/google goggles. People inside google/microsoft are racing at this.

This really depends on what you mean by "solved." Kinect is a hell of a long way from what you would need to avoid collisions in a 3D space. Kinect basically just has to deal with the outlines of objects at a relatively narrow set of distances. When your sensor is moving in three dimensions and you are trying to track an object that is also moving in three dimensions it gets a heck of a lot harder, even if you are using radar (which gives you range). If you're trying to figure out range based on the apparent size of an object of unknown actual size, it gets even harder.

>I'm sorry, i have to disagree with your pessimistic attitude on this.

I'm actually quite optimistic that it will happen, just not for many years yet.

>With regard to fire - fire can kill human pilots too. With sensors, you can create a multiply redundant system - put in 20 extra sensors. With humans it's not possible.

If your engine is out on a wing and it catches fire, the fire sensors will tell you so, and shortly thereafter they will most likely be destroyed and tell you nothing further. An engine on fire out on the wing is not going to burn up the pilot. The pilot can look out the window and quickly and easily assess the condition of the engine and the wing: did it burn out, or is it raging out of control, or maybe there are subtle signs that indicate something in-between? Maybe with enough fire sensors scattered all over the plane a computer could make a similar assessment, but you're talking about a lot of extra money and weight, and you still have the problem that your sensors are going to burn up shortly after going off and then you have no idea if the fire has gone away or if it has just stopped spreading. Someday maybe you can give the computer a camera to "look" at the wing to make the same kind of assessment that a human pilot can make, but that is a very long way off.

[marshallp]

You have two main pessimisms

- computer vision problems (e.g. plane catches fire, how do you tell how much fire etc.)

- tracking other objects in 3d while in moving in 3d at high speed

The question is - can humans do this? If yes, computers can do it eventually. The only question is how long away is this. What we know from machine learning, is that data is important. If you can gather enough data you can do anything. So really, your problems are a question of data collection. It is not a technically difficult problem. (By the way, one of the problems you have in aerospace is, is that you're control theory heavy rather than pro-ai, which means you end up not being able to solve the difficult problems.)

Also, the eu 80s project ended around early 90s, and the grandle challenge win was only 15 years later, not 30. Had the challenge been tried 5 years earlier, it would have worked. The algorithms and hardware was already sufficient.

[lucasjung]

I hardly think it's pessimistic to say, "this problem is really hard, and it's going to take years to solve." I'm confident that they will be solved, which some of my peers might even consider a naively optimistic attitude.

>The question is - can humans do this? If yes, computers can do it eventually.

This logic is deeply, deeply flawed. I happen to believe that computers can eventually perform the tasks under discussion, but "humans can do it" is not one of the reasons why I believe that.

>So really, your problems are a question of data collection. It is not a technically difficult problem.

When it comes to tracking airborne targets with airborne radar, data collection actually is a technically difficult problem. The combination of waveform, antenna design, transmitter design, receiver design, tracker design, etc. present a set of engineering tradeoffs in effective range, range resolution, azimuth resolution, weight, size, flase positive and false negative rates on radar returns, and other performance characteristics. Even the very best airborne radars provide data which is limited, especially in terms of accuracy and precision.

>By the way, one of the problems you have in aerospace is, is that you're control theory heavy rather than pro-ai, which means you end up not being able to solve the difficult problems.

A little more about my background: my BS is in EE, with a specialization in microcomputer interfacing (I took a lot of CS classes). In grad school, my stability and control prof had actually done some pioneering work in incorporating non-linear logic into stability and control systems (don't try this at home, kids). In addition to stability and control, my other focus for my MS was avionics. The prof who taught most of my avionics classes was actually from the CS department (his undergraduate background was EE, with a specialization in radar). One of the things that kind of surprised me about aero, having come from EE, was how broad the discipline is. Before going back to school for aero, I thought that getting a degree in aeronautical engineering would be primarily about aerodynamics, with a smattering of other stuff. Instead, I discovered that everyone gets a little bit of everything (aerodynamics, propulsion, structures, stability and control, avionics), and then specializes in one or two particular areas. By the PhD level, people who have specialized in areas other than aerodynamics have largely forgotten most of what they learned about it as undergrads. It's an incredibly heterogeneous field: propulsion and structures guys have more in common with MechEs than with other AeroEs; avionics and stability and control guys have more in common with EEs than with other AeroEs; etc. So to characterize the discipline, or any individual within it, as "control theory heavy rather than pro-ai," displays a deep misunderstanding of the character of the community. I guarantee you that there are plenty of AI experts working in the aero field.

>Also, the eu 80s project ended around early 90s, and the grandle challenge win was only 15 years later, not 30. Had the challenge been tried 5 years earlier, it would have worked. The algorithms and hardware was already sufficient.

This just reinforces my broader point: success came not as a result of some innovative genius applying a novel new approach but rather because the technology had matured--over the course of several years--to the point where success had become not only possible but likely. Radar tracking is currently experiencing big advances for that same reason. The theory behind Space-Time Adaptive Processing (STAP) has been around for decades, but the available technology has not been up to the task of implementing it effectively. In the past we've resorted to less effective tracking methods such as MTI, but in the last decade or so the technology has finally made STAP reasonable to implement.

[marshallp]

It seems like you are not well versed in ai. There was a debate about whether "hand engineering" vs "dumb simple algo's" would get results. Dumb algo's won. Your mentioning of MTI and STAP, and how difficult radar design is etc. etc., makes me think you aero guys are still in hand-engineering land.

I will offer another example. Why did the aerospace dudes not be able to autonomously fly a helicopter. in 2004, andrew ng decided to tackle this. He completely ignored any previous work, just using a dumb algo (reinforcement learning) and laptop managed to get amazing autonomous performance out it. Why was it him (ai researcher), and not people from the field of flying.

[lucasjung]

>It seems like you are not well versed in ai.

Not especially. I audited a course as an undergrad, and hardly remember anything from it now, but I have a layman's understanding of the basics.

>There was a debate about whether "hand engineering" vs "dumb simple algo's" would get results. Dumb algo's won.

There are design tasks for which "algorithms" are better suited, and there are design tasks where experienced human engineers still do far, far better. The statement "Dumb algo's won" is certainly true for some applications, but not all.

>Your mentioning of MTI and STAP, and how difficult radar design is etc. etc.,

Now it's my turn: you clearly are not well versed in aviation or in radar principles. Not every problem can be magically solved by throwing AI at it. Radar theory is well established, and the equations are well known. Unfortunately, they are hard equations to solve: determining the location of an object using radar involves some complicated math with a lot of variables, and the only way to solve those equations is to chew your way through them. You can simplify them, but then you have to accept increased errors from the terms you throw out.

Where "algorithms" come into play is tracking, and depending on how you define AI, radar engineers have been using AI since the invention of the first automated tracker. Even the very best automated trackers in existance today are not nearly as good as an experienced operator looking at raw returns. Someday that will probably change, but that day is still many years away.

>...makes me think you aero guys are still in hand-engineering land.

As I said before, you're making a huge mistake by lumping "you aero guys" into a single group. "Aeronautical engineering" is really "every other kind of engineering, applied to aviation." When I was in grad school, one of my buddies' thesis was pure AI: he developed a learning algorithm for choosing the optimum path for a jet to taxi around a crowded flight deck, using DGPS as the only position source. Another guy combined machine learning with CFD in an attempt to design better supersonic lifting surfaces (the results were not good, but his thesis was still a "success" in the sense that he expanded human knowledge and the general concept showed promise). There are some applications where AI is the way to go, and there are some applications where what you derisively call "hand engineering" is infinitely superior.

>I will offer another example. Why did the aerospace dudes not be able to autonomously fly a helicopter. in 2004, andrew ng decided to tackle this. He completely ignored any previous work, just using a dumb algo (reinforcement learning) and laptop managed to get amazing autonomous performance out it. Why was it him (ai researcher), and not people from the field of flying.

Time for a history lesson:

The Navy deployed the first effective autonomous rotorcraft in the 1960s: http://en.wikipedia.org/wiki/Drone_Anti-Submarine_Helicopter

So the "aerospace dudes" were "able to autonomously fly a helicopter" before Andrew Ng was born, and they did it using "hand engineering."

Not autonomous enough for you? Firescout flew autonomously four years before Andrew Ng flew his helicotper autonomously: http://en.wikipedia.org/wiki/Firescout

Firescout was also developed by "people from the field of flying."

[marshallp]

There's quite a difference between firescout and ng's helicopter. The latter gives superhuman performance (there are videos on his homepage, andrew ng stanford). Anyway, i don't know what algo firescout uses, it might well be ai behind the scenes proving my general point.

Your point about complicated mathematical equations lies at the root problem of you "unified engineering" guys. modern ai (machine learning) is where you give up on the assumption that you (puny human) can impart "wisdom" to your system. You simply throw a random set of equations (a neural network) that are large enough/not too large (overfitting) to capture physical reality. Getting the errors low is a matter of getting enough data and experimentally adjusting the size of your nnet.

Yes, there mught be grad students and profs trying ai to solve aero problems, however, if enough resources are not devoted, they will not yield good enough results. For example, spend a billion dollars (gathering data/computation) to solve your radar problem. A billion dollars in your field is pocket change.

[lucasjung]

>There's quite a difference between firescout and ng's helicopter. The latter gives superhuman performance (there are videos on his homepage, andrew ng stanford).

You have absolutely no idea what the performance and handling characteristics of Firescout are; in fact, it is apparent that you lack the domain knowledge to understand their meaning even if they were presented to you. Nevertheless, you assert without hesitation that Andrew Ng's helicopter is superior. This is the epitome of fanboyism.

>Your point about complicated mathematical equations lies at the root problem of you "unified engineering" guys.

You clearly don't know what you're talking about here. Go read Skolnik's Radar Handbook, then try to tell me with a straight face that random processes are going to derive those equations for you, and somehow magically come up with a way to sidestep the basic reality that they have to be solved.

>modern ai (machine learning) is where you give up on the assumption that you (puny human) can impart "wisdom" to your system. You simply throw a random set of equations (a neural network) that are large enough/not too large (overfitting) to capture physical reality. Getting the errors low is a matter of getting enough data and experimentally adjusting the size of your nnet. Yes, there mught be grad students and profs trying ai to solve aero problems, however, if enough resources are not devoted, they will not yield good enough results. For example, spend a billion dollars (gathering data/computation) to solve your radar problem. A billion dollars in your field is pocket change.

In order to use a tool effectively, you have to understand both it's capabilities and it's limitations. Even though you have indicated that you are an expert on machine learning and I have admitted that I am not, it is now abundantly clear to me that you have absolutely no understanding of the limitations of machine learning.

If only every problem could be solved optimally by simply throwing enough data and a big enough net at it. Unfortunately, that's not how the real world works.

One problem is that machine learning algorithms often converge on local maximums that are far less optimal than is possible. The guy who worked on the taxi routes had enormous issues with this, and only after extensive tweaking was he able to come up with solutions that were on par with human path-choosing.

An even bigger problem stems from the fact that a machine learning algorithm is only as good as the model it works in. I mentioned that the guy working on lifting surfaces in CFD did not get great results. His problem was that his algorithms pretty much always found the places where the CFD models diverged from reality: they would find the optimum shapes for the model they were working within, but those shapes always performed terribly in the wind tunnel because the algorithm was finding optimums at points where the model diverged significantly from reality. You can't solve this problem with "better models," because every model diverges from reality. If a model doesn't diverge from reality, it's no longer a model, it's reality. Where he really impressed his review board was when he detailed a follow-on experiment of using this phenomenon to develop better CFD models, within which human engineers would be able to come up with better designs.

Finally, a billion dollars is not "pocket change" in any field. Even if someone had a spare billion dollars laying around to fund R&D for radar tracking, the opportunity costs of blowing it on a machine learning experiment, instead of using to fund experienced engineers working from proven principles, would be unacceptably high.