FA+
Shop
Artifical Intelligence Ethics
9 years ago
General
I've got an interesting question for you.
As some of you may know, I like to play a great little air traffic control simulation called ATC Windows. The fact that I play an air traffic control simulation to unwind might tell you something about the way my days usually go, but I digress.
There are many different commands you can issue pilots in the simulation, one of which is an altitude crossing, such as "cross MAZIE at 11000". This tells the pilot that they can (in this example) descend at their discretion as long as they cross the MAZIE fix at 11000 feet altitude.
These days, most commercial airlines are very conscious about fuel costs and pilots are encouraged to stay at altitude for as long as possible before descending. However in ATC Windows, it is no secret that the pilot AI is programmed to look for aircraft crossing their path and if they've been issued a descent at their discretion, they will descend to purposefully cause a loss of separation with that crossing aircraft.
I'm torn about this. Obviously, pilots in the real world don't watch their TCAS for other aircraft and do their best to mess with the air traffic controllers. In the above example, on approach to Philadelphia, aircraft enter New York Center Sector 97 at the north border of the sector with the MAZIE fix almost 100 miles down their route. When a pilot first checks in, if you told them to cross MAZIE at 11000 feet (they ordinarily enter the sector at 17000), I think it would be reasonable to expect them to stay at altitude until only 10 or 20 miles before MAZIE, depending on the aircraft and its performance profile.
On the other hand, just because pilots like to stay at altitude doesn't mean they will. As an air traffic controller, if you issue a descent at pilot's discretion, you should be ready for them to descend at any time. Their path should be clear of all other aircraft that could cause a conflict. There are several flight paths that cross under the Philadelphia arrival stream, and if the controller decides to descend LaGuardia jet arrivals early, they cross under the Philly arrivals as well. All of these are conflict sources.
The simulation author's reasoning is that the AI behaves this way to break you of bad habits. And while I can't disagree, it still irks me to know that the AI is programmed to act in a way no real pilot would ever behave. If the goal is to simulate the real world as closely as possible, the AI should behave like pilots in the real world.
What do you think?
As some of you may know, I like to play a great little air traffic control simulation called ATC Windows. The fact that I play an air traffic control simulation to unwind might tell you something about the way my days usually go, but I digress.
There are many different commands you can issue pilots in the simulation, one of which is an altitude crossing, such as "cross MAZIE at 11000". This tells the pilot that they can (in this example) descend at their discretion as long as they cross the MAZIE fix at 11000 feet altitude.
These days, most commercial airlines are very conscious about fuel costs and pilots are encouraged to stay at altitude for as long as possible before descending. However in ATC Windows, it is no secret that the pilot AI is programmed to look for aircraft crossing their path and if they've been issued a descent at their discretion, they will descend to purposefully cause a loss of separation with that crossing aircraft.
I'm torn about this. Obviously, pilots in the real world don't watch their TCAS for other aircraft and do their best to mess with the air traffic controllers. In the above example, on approach to Philadelphia, aircraft enter New York Center Sector 97 at the north border of the sector with the MAZIE fix almost 100 miles down their route. When a pilot first checks in, if you told them to cross MAZIE at 11000 feet (they ordinarily enter the sector at 17000), I think it would be reasonable to expect them to stay at altitude until only 10 or 20 miles before MAZIE, depending on the aircraft and its performance profile.
On the other hand, just because pilots like to stay at altitude doesn't mean they will. As an air traffic controller, if you issue a descent at pilot's discretion, you should be ready for them to descend at any time. Their path should be clear of all other aircraft that could cause a conflict. There are several flight paths that cross under the Philadelphia arrival stream, and if the controller decides to descend LaGuardia jet arrivals early, they cross under the Philly arrivals as well. All of these are conflict sources.
The simulation author's reasoning is that the AI behaves this way to break you of bad habits. And while I can't disagree, it still irks me to know that the AI is programmed to act in a way no real pilot would ever behave. If the goal is to simulate the real world as closely as possible, the AI should behave like pilots in the real world.
What do you think?
- If it's a *game*, then it's a nice make-things-more-complex addition.
- if it's a *teaching tool* then it's definitely a worthwhile change.
- If it's a *research* or *planning tool*, then it's a bug.
It's more than a game. The character-graphics-based ATC game I used to play on my uncle's XT machine (yes, this was long ago) was a game. Simple keypresses to command aircraft, a square sector with two "airports" and different entrance/exit locations around the border.
In ATC Windows, pilots speak to you in proper phraseology ("Delta two-nine-six heavy with you at seven-thousand, six-hundred cleared one-zero-thousand") and you are expected to talk back to them in proper phraseology. The sectors are modeled after real-world sectors, including different ceilings and floors (ZNY97 has an altitude "shelf" on the west half). You deal with real radio frequencies for adjoining sectors. You accept handoffs from other sectors and hand off to them. You type near-accurate commands into the display computer to alter the display. There is weather. There is turbulence, pilots ask for ride reports (and you're expected to reply) and if the turbulence gets too rough they will request a different altitude. There are aircraft that come into your sector that are not RVSM certified, requiring extra vertical spacing with other aircraft.
So if you're going to go so far in making a real-world accurate simulation of air traffic control... why tweak it to diverge from that model on purpose? And once that information gets out, how can you trust other aspects of it?
In Chicago Center sector 82, aircraft going to Detroit must be descended to flight-level 240 before handing them off to the next sector. If two aircraft come into the sector close together along that route and one of them is going to DTW, it seems like 90% of the time the aircraft going to DTW (the one needing to descend) is above the one that's flying through. Similarly in New York Center sector 97, if you have a departure out of Allentown climbing toward SEG, you can almost guarantee you'll get an inbound prop to La Guadia, which enters over SEG, flies close to the same route as the ABE departure and needs to be descended.
I agree, if it's a teaching tool, it's good to have the extra challenge for practice. ATC Windows has a career mode that includes training and certification. It keeps performance records for your sessions. And if you really screw up (a "loss of separation") you're usually de-certified in a sector and need to do manditory training before attempting to certify again. So maybe the altered AI model is okay when training. But when you're supposed to be working a sector... shouldn't it be like the real world?
For that manner you should create the occasional pilot who just does something really stupid and in violation of the rules - because human error happens, no matter how well the pilots are trained.
So if there's some random variance of when a pilot descends when given a crossing, I don't mind. I did clear them to descend at any time after all. I think what nags me is that I know I'm going to mess up and issue a crossing when it's not completely clear. I've learned I can't rely on routinely using crossing commands to stay clear of traffic, but if I slip up I'd like a chance that it doesn't end in getting written up. Kinda like the real world.
But it could be modified to execute the descent in some kind of window, then it would more closely a real world scenario. But like others said, its meant more of a training and practice tool, which probably means its *intended* to keep you on your toes and monitoring everything.
I thought about it a bit and came up with a parallel in racing. What if I wrote a racing simulator where as soon as you put all four wheels off the track, your car spun out? You're not supposed to put all four wheels off when racing and you should try to avoid it. It does happen and you could spin out if you do. But you don't spin out every time you put four wheels off track.
Or how about if I wrote it so when you bump another car, you're immediately disqualified? Despite what was said in "Days of Thunder", rubbin' is not racing and is frowned on by the stewards. It does happen though, sometimes on purpose. But every little bump and rub shouldn't end in immediate disqualification. That's not real-world accurate.