Comment by giantrobot
3 months ago
> I am completely naive about the challenges presented.
The problem isn't collision detection or predicting movement. They're not a bunch of particles on simple ballistic trajectories. They're powered objects traveling in a turbulent and difficult to predict medium. In emergency conditions they can turn from a powered vehicle to an unpowered one. They can need to land immediately when flight worthiness changes in flight. A situation on the ground can make landings unsafe or impossible and an aircraft needs to diverted disrupting traffic at another airport.
Automating ATC works until one or more exceptional conditions arises. Then it's completely unsuitable and everyone from pilots to ATC need to work against the happy path automation to keep people alive.
Known exceptional conditions can all be modeled and simulated preemptively. Like on each position update, for each plane, what does the overall situation look like if it needs to all of a sudden make an emergency landing.
(and just to be clear, no I'm not talking the "AI" genie but rather straightforward search algorithms that enforce the needed invariants)
Aircraft work within an envelope. You can model what a max speed/min speed max/climb min climb/ trajectory is and work that in. It's a solvable problem. Aircraft can't go from 200-0 knots or 0-200 knots instantly etc
200-0 happens regularly. especially with GA aircraft. The plane will be scattered across a field most likely.
This is a general objection to AI responding to real world events in general : "What if something unexpected happens?" It comes up in self driving as well. Things like "What if something suddenly appears in the middle of the road" or "Can it drive in snow conditions with zero visibility?
My question is, how do you know that in general human beings respond better to unexpected or very complex / difficult situations than an automated system would? Yes, human beings can improvise, but automated systems can have reaction times more than an order of magnitude faster than that of even the quickest humans.
I'd like to see some statistics on the opposing hypothesis : How good are humans, really, when encountering unexpected situations? Do they compare better with automated systems in general?
Here's a competing hypothesis: An automated system can incorporate training data based on every recorded incident that has ever happened. Unless a situation is so unexpected that it has literally never happened in the history of aviation, an AI system can have an example of how to handle that scenario. Is it really true that the average human operator would beat this system in safety and reliability? How many humans know how to respond to every rare situation that has ever happened? It's at least possible that the AI does better on average.
In theory, everything works. In practice, we can't even master automated driving, on two dimensional streets with painted lanes, relatively slow speeds, and cars that can just stop in case a decision could not be made. If we can't make this happen, how do you expecct the same with higher speeds, an additional dimension, planes with radio-only (no additional telemetry) and pilots with heavy accents?
>Unless a situation is so unexpected that it has literally never happened in the history of aviation,
I would say this is actually the most likely scenario for an edge case. The sheer number of variables make it unlikely that the same unexpected event would happen twice.
In an emergency situation the combination of, the emergency, ground conditions, weather, visibility, instrumentation functionality, and surrounding aircraft is most frequently going to be unique.
> I'd like to see some statistics on the opposing hypothesis : How good are humans, really, when encountering unexpected situations? Do they compare better with automated systems in general?
This is already out there. You can go research how Airbus and their automation works in practice.
You can also listen to air traffic control recordings to get an idea of what types of emergencies exist and how often they happen. I'm sure the FAA has records you can look at. :)
Now that apply that to something 3 orders of magnitude more complex.
Air traffic control is almost 100 years old now. Unusual things happen, yes... but unpredictable ones do not. Ever. No conceivable emergency in ATC cannot be handled by a machine following a procedure, if it could be handled by people following the same procedure.
Just another perplexing case of humans insisting on doing a robot's job for no good reason.