People don't realize how powerful a basic computer really is.  Especially when it's not doing crap like painting pretty pictures on a screen so a human can have a nice friendly interface.

I'm sure a military simulator is several orders of magnitude above anything a PC combat simulator ever thought of being.  I'm sure it's very accurate to real life with the biggest difference being the simulation of G forces on "Human" component, or lack there of.

I've heard it said that FAA regs are the only reason they aren't.

Why do they have cockpits?

I'm just a grunt, but I would think the physiological limits of a human pilot would be a distinct disadvantage versus AI.

Humans GLOC at what, 8 or 9G?  Computers don't.

I know there's a lot of "stuff" (i.e. armor, structure, ejection seats, etc.) in aircraft just to make them more survivable for the occupant(s) when hit and/or crashworthy.

Granted, one would still have to protect (armor) the AI, but without the need to protect the crew from injury it would see one would free up a lot of "resources."

Already in use now- look up MUM-T

Automation isnt going to take most human jobs and lead to an end of society as we have known it.  Nope everythings cool.

I'll keep this very top level. I have also worked extensively with these systems. Building, collecting empirical data, making modules, etc etc.

6DOFs are built on a metric fuckton of empirical data - non specific to whatever is going to be simulated inside of the 6DOFs. Additionally, many many sensors, sensor components, and sensor systems are part of 6DOF. Again, non-specific to whatever will be flying in there. Weather, atmosphere, gravity, how environmental conditions effect what sensors see, etc etc are all in there already. The 6DOF only cares about how something interacts with it's environment. The 6DOF doesn't control what that thing "inside of it" is doing.

For whatever is going to be put into the 6DOF - the parameters of the vehicle, things like aero, autopilot, ofp, what sensors and where, etc etc - those have their own module. We'll call that the "Robot Overlord" module. The 6DOF program will feed exactly what the sensors are seeing to the Robot Overlord module. The 6DOF program controls all the physical and EM interactions with the Robot Overlord. It does not control the Robot Overlord itself.

The AI is simply a component of the Robot Overlord module, all of the empirical data that its sensors would be “seeing” are fed to it. There is no need for the AI to simulate the weather, because that’s not it's job. That's not what it would do out in the field. The AI only cares about controlling itself, based on what it's sensors are telling it. It is the 6DOF program's job to simulate the weather, environment, what those sensors can/cant see, etc. So... if the AI commands an aileron deflection – the aileron module is feed the new forces on itself as it interacts with the atmosphere and wind and weather and other aero forces from the vehicle itself from the 6DOF – this translates into some positional/voltage/amperage/resistance/whatever data that is fed back to the AI.

This is done for everything on the vehicle. For all intents and purposes the AI – or operational flight parameters – of the vehicle inside of the 3DOF, doesn’t know it’s in a simulator. So yes, the AI would “see” video feeds or IR or MMW or whatever other sensor/feedback it would see in real life.

That is how they work in general.


-eta- Just because it is run on a standard computer, doesn't really mean much. You typically need the extra horsepower of a compute cluster for monte-carloing simulation runs. A single run, wont choke a modern computer. Running 10 thousand runs at once? Your computer is just gonna say fuck you and go surf porn instead.  Of course this is asterixed - depends on what all is in your sim... also edited abit for clarity.

It's almost the opposite.  Off the cuff, I would joke that more effort goes into avionics support/survivability than pilot support/survivability.  And, I think a review of charge numbers and man hours might show that it's not a joke.

There isn't any armor protecting the pilot, and avionics would never be armored, even if it was a UAV.  The defense is redundant avionics units, which protects against the more likely system failures, as well as the less likely battle damage.  For something like a helicopter that could be expected to take small arms ground fire, maybe there is some armor there, but certainly not in a fighter platform.

In a computerized simulator, the computer knows everything that is happening.

Put that computer in a real aircraft even with 360x360 degree cameras, and the computer will fail miserably unable to process what is happening and what is the priority this second.

Air combat is like playing chess, you plan ahead, except there are not a finite number of moves to analyze.

That and the constant instruction from ATC

That and the need to sell tickets.

Nobody is going to get on an aircraft that is completely automated.

Yeah we will.  Plenty of people will have no problem getting on a completely automated flight.  At least you know the pilot won't decide to suicide out into the side of a mountain.



 

When autonomous carrier landings are so routine that nobody cares anymore, I'll start to contemplate this as a possible future problem.

I have to disagree that a computer will always be worse than a human at "processing the situation".  In fact, I would say it is the exact opposite.  Put a human in a real aircraft, and most of the time it will fail to process everything happening around it, and then act accordingly.  The potential is there for the computer pilot to massively exceed the human "bandwidth" capacity.  A software pilot can watch and react to everything around it in real time, while the human is limited to looking at and concentrating on a single task at a time.  As the situation grows more complex, the computer has the potential to gain an ever increasing advantage over humans.

Your chess analogy is also interesting.  Eventually, we will have a swarm AI controlling all "drones" in the theater.  At that point, it will be like a group of human players all individually in control of a single white piece trying to play against a computer controlling all black pieces simultaneously; checkmate.

ACLS makes that possible today (and long in the past), but the meatbags at the controls don't like to let the computer do that for them.

I love that this post came just after the post a few responses up.

 

But I want sex bots now.

From the original article, at the very top in large print.

That's one way to solve the pilot shortage.

You are joking right?

The materials the plane is built out of has limitations too. But why is everyone focusing on the acting when the OODA loop is a thing and the 35 is very good at the OO?

looking forward to a day we can transcend humanity with cybernetic.?

Truly autonomous computer controlled carrier landings have happened a handful of times under extremely controlled conditions.  They're still something of a big deal.

When that is the norm rather than the exception, we can begin the discussion (IMO).

As long as ego driven stick actuators are the norm, we will continue to pretend that people are better than computers at simple problems...

https://news.usni.org/2013/11/22/navy-completes-initial-development-new-carrier-landing-system

Not for the first time, a real conversation shows that the shenanigans of GD overlay some real minds and experience.



I like this thread.  Some good issues have been thrown onto the table.



There's going to be some relevant activity going public here:  http://www.ecai2016.org/



Not sure how much I want to say beyond that, at this time.  Let me drink my coffee and ponder that.

I need to get off my ass and finish acouple of papers I have. My org pays for my "vacation" whenever I publish and present. I've leveraged that many times.

Looking at the accepted ECAI papers and the Topics of Interest... I could have been getting ready for my Holland vacation right now



GD is weird. I've tried many times to get some "pure" tech discussion going on various topics. Shit goes nowhere. You basically have to have just enough trolling for the thread to survive long enough the right eyes eventually see it.  Window lick for a little bit, then boom start the real thread.

You would not trust automation if you have seen what I have seen while on automation.

The automation would have to be light years ahead of what we have today with much much more backup and cross checking.

I will never get on an aircraft without a pilot.

Period.

Not talking point to point flying. My reference is air to air combat.

It is too fluid and too complex. I don't believe a computer could capture all the needed info real time. The problem is the sensors, not the computing power.

But could it beat a naval fighter pilot? Probs not.

Not today, but tomorrow?





Today, performance parameters are limited by the meat inside, and control systems have to interoperate with multiple threat/target ID-acquisition-targeting systems.  Executive decisionmaking is in the hands (and brain) of the pilot.





But machine-based complex decisionmaking is getting better and faster.  And a fair bit of this stuff doesn't need to be "AI," but rather just heuristics-based.  Attack UAVs with human oversight, or maybe general outcome selection (but not active hands-on control), are down the road, yes?





 

What most folks refer to as AI, is just an aggregate of various heuristic functions. But yes, once you characterize the decision space; solutions can be reached incredibly rapidly because at that point it's essentially just filters and array pointers. What can seem like incredibly complex decisions can arrive at the optimum solution all but instantly.  For example PCA a series of different observation arrays, load/bias the principles, throw it into a bayesian tree. Bingo-Bango you just made an incredibly high ordered multidimensional problem an all but instant linear problem. One that is very easy to bias or update - with little over-fitment.

The "brains" actually take very little horsepower if designed properly. The crap that still takes time is the sensor conditioning. Sans implementing filter specific FPGAs - Fourier Transforms and Convolutions is where cycles get eaten up.  

These already exist and are in field, and have been for awhile.

The US will never have man completely out of the loop on military systems, simply for liability/safety issues. Even if it's simply an abort option, it will always be there. As "smart" as you make your system - a short that wipes out a code block in your NVflash and you're fucked. You will always need some meatbag control functionality as a failsafe.

Having said that, the platforms I developed a decade ago were designed and operated completely autonomously. Everything.  We could over-ride if needed, but never did.

Really meaty stuff.  Thanks!
You kinda cut to the heuristics design chase, in terms of pure performance.  For me (this thread is nominally about AI), Bayesian and math/statistics approaches that catch statistical clusters, make consequent parameter changes, and then feed those changes back into the system in an iterative way to create adaptive systems aren't AI at all.  Those approaches go directly to performant subsystems, yes.  They're effective and comprehensible, yes.  They're easily coded in available silicon, yes.  But taking a step up, into the executive decisionmaking realm, with a solid cognitive architecture that constrains behavior in a process that can handle fuzzy data and fuzzy logic, that finds some balance somehow (magnets!) among all the subsystems providing data or decisionmaking functionality, is where AI starts to mean something for me.  Humans do this stuff all day every day, but it is fricking hard to model a process that we only dimly understand.  And by that I refer to human cognition.



[ETA: and quantification of human values]
With autonomous weapons systems, things like ethics and morality suddenly have a place front and center in engineering.  But how do those things get put in?  A partitioned engineering approach generally means that a concern is out of scope for someone's responsibility.  Our good friends at [redacted] are very, very good at partitioning and abstracting real-world problems into acceptable and little puzzles that, when solved separately without context, can get put back into a system that is perhaps-- well, let's not go there.
It's the general "Who watches the watchers" problem.  We're going to have to do this with laws, which, except for the laws of physics, are merely advisory.
 

True to GD standard, I only just now read the link in OP.   Actually, I read the research paper. I can't stand "science/tech" article reporting. They did it exactly how I figured they did, cluster training/tuning on a faster computer - then characterization on the Pi.



This kinda gets into, for me at least, a murkey area. At what point do you call something an AI? Many thing I work on, average joe 6 pack is going to think it's an AI. To me it's just characterized state/decision spaces, biased or weighted for a desired optimal solution. Yeah, there is a fuck ton that goes into it, not down playing that. Optimizing a limited decision space simply doesnt pass my personal filter for AI. OP "AI" for example, to me that's just an OFP with branches for positional superiority and threat analysis optimizations. When broken down, it's coordinate optimization based on risk. A take on the classic predator/prey, but using a bayesian tree with fuzzy state/space so the "old grey beards" can tweek the gains easier. But I guess because I see the sausage getting made - I've got a different viewpoint on it.  

Maybe I should come up with some criteria, at least for myself, as to what I think starts constituting AI. The Turning test sure as shit isnt a good scoring device. Maybe a ratio of the initial decision space vs ability to adjust for something truely outside of it?  Yep, that it. Pattenting that shit, the outofstep ratio. It even works with my name.  

I personally think we are decades away from anything that will have a good outofstep ratio. My concern for a system being able to do something well outside initial training is pretty much zero.

A poorly trained tree controlling a critical system? Now that's a different story.



In military acquisitions, it will be in the RFP. The project wont even get off the ground unless it is pretty explicitly in there from the get go. The liability of not having a human give the final go ahead for terminal effects, is staggering. Trust me, Ive seen 1st hand for many years how people react to even the potential of having a weapon system they cant take full control of or that doesn't do exactly what they commanded it to do.  

Now... Industry and other countries arent limited in that way. Control of a critical/dangerous system, without having a man in the loop. That could cause some problems in a hurry.



I dont think I've ever met an engineer that hasnt read I, Robot.

X drones already fly completely autonomous missions. Predators already target and fire missiles. I'd be stunned if these three things are not already a thing somewhere.

Uh... why not?

I would be shocked if it couldn't, with the current state of the tech.

Is this before or after their on-board computer systems are compromised?

Yeah. That's it.

So, you're comfortable with an A300 tooling over your house controlled by a data link whose cryptology will be known to every country on the planet?

We don't have a pilot shortage, so there is that flaw in the analysis.

I can think of two or three Airbus accidents attributable to poor meat computing. This is the current state of the art in aircraft automation.

Yup. We (humans) are programmed to move forward as fast as we are able (and at times, much faster than we should). It's part of the deal. This runaway train isn't going to get stopped until it flies off the tracks and violently smashes into the riverbed.

We actually have a science and math subforum, and as long as its peripherally aviation related, put it in the aviation forum.

Math and science subforum gets about 0.000001% of the eyes GD does. The last time I went in there, I spent like 3 pages trying to explain what a derivative/integration was and that you cant just take the square root of some random acceleration value and get original position and that you cant just take the cube root of a volume to arrive back at the original length of one of the dimensions. You cant rearrange 10 ft^2 with 3 ft^3 and get 13 ft^5...  I mean jesus... basic unit manipulation...    After that jumble of pure fuck, I'm done there. I'll take some window licking out here, if it means people that know wtf they are talking about eventually see the thread. Besides, some window licking adds to the flavor.

I meant the looming pilot shortage that is supposed to be coming. I read a doom and gloom article about it. It could be complete bs, as I clearly haven't a clue about pilot manning. Nor do I particularly care.

I've enjoyed the thread.  Some surprising depth, and in areas that help me learn.



If anyone is going to attend one of the following, hit me up with an IM.



IEEE World Congress on Computational Intelligence

European Conference on AI

International Conference on Soft Computing Models in Industrial and Environmental Applications/Computational Intelligence in Security for Information Systems

Pay, benefits, work rules and retirement have all taken massive cuts in the last 15 years and it no longer attracts people willing to invest years and over $100,000 to get a $30,000 a year job spending most of the month away from their families.  That is the pilot shortage. It is people that would have liked to do it say no thanks and do something else.