More DFH news

DaWolf85 said:
That's not really how computers work. Getting them to do the basics is relatively easy and quick. You can refine that into perfect handling of most jobs, 75-90% of the time, in relatively short order. But replacing humans for the remaining part is not just much harder, but actually currently an insurmountable task.

Theoretically, it would be possible to train a neural network to do much of the work of dispatching, potentially even that last 10-25% (at great expense, mind you). But the danger of neural networks is that they are not auditable. You can't really know exactly how they're breaking the problem down, or what little issues might trip them up. In short, they're not actually suitable for safety-critical applications. There will always need to be humans watching the output and correcting it, no matter how good the automation is. And given that humans are bad at managing automation in that way, especially weird, inconsistent failures of automation as you would see from this 'experiment', that means it's safer to have the human do it from the start.
Click to expand...

If an automated system built through ML can spit out flight plans accurately 75-90% of the time the airline just needs the human to check over it and correct it 10-15% of the time. On a 50 flight desk you're correcting 5-12 flights. Those are realistic figures today, after the system has went through a lot of training, but it's a big field making big advancements so it isn't unlikely that in 15-20 years a system like this could be accurate over 90% of the time. In that case you're correcting 1-5 flights on a 50 flight desk. Why would they pay anybody six figures to do that and why would they hire so many people like they do today? Maybe they'd concede to keeping the wages through union pressures but they won't have a need to hire as many people as they do today so the dispatching industry will just shrink. If automation causes an airline to only need to hire 5 dispatchers instead of 10, then automation has caused dispatching to lose 5 jobs. In some ways it is similar to the fall of the flight engineer position.
 
flynryan692 said:
If an automated system built through ML can spit out flight plans accurately 75-90% of the time the airline just needs the human to check over it and correct it 10-15% of the time. On a 50 flight desk you're correcting 5-12 flights. Those are realistic figures today, after the system has went through a lot of training, but it's a big field making big advancements so it isn't unlikely that in 15-20 years a system like this could be accurate over 90% of the time. In that case you're correcting 1-5 flights on a 50 flight desk. Why would they pay anybody six figures to do that and why would they hire so many people like they do today? Maybe they'd concede to keeping the wages through union pressures but they won't have a need to hire as many people as they do today so the dispatching industry will just shrink. If automation causes an airline to only need to hire 5 dispatchers instead of 10, then automation has caused dispatching to lose 5 jobs. In some ways it is similar to the fall of the flight engineer position.
Click to expand...
The issue I have with your statement here is not that flight plans are inaccurate but that there are many factors that automation does not look at that a dispatcher would so even if a flight plan is accurate strictly from a 121 standpoint that doesn't mean it is a flight plan that I would sign my name to as a dispatcher. A great example is on transatlantic flights. Most automated flight planning systems don't look at Turbulence that closely and really don't take into account pilot reports or what the turbulence will do in the future, which are things almost any competent dispatcher will consider. So on a EWR-LHR flight for example, an automated system may plan a flight on NAT Track Z because that is the best wind route. However, a dispatcher will look at Track Z and see that there will be moderate turbulence between 40 and 20W. However if that flight were planned on NAT U, there would be much less turbulence and a better ride for the customers. I as a dispatcher would then plan it on NAT U. There was nothing inaccurate about the NAT Z flight plan, but NAT U was obviously a better choice.

And you don't take into account the stuff that happens after a flight is released. You could have ten flights from the northeast to Florida, and after they are checked over and sent ATC closes the AR Routes, and now all of a sudden you have to process reroutes for ten flights at once. This is why we are still needed. One other example I will give is I recently had a flight that couldnt land in ORD for windshear and my first instinct was to divert to STL. However, STL said they could not fuel the airplane, so I quickly had to change the diversion to DTW who could fuel the airplane. As far as the system goes, STL was legal. It's the things like this that lead me to believe that Automation will never eliminate our jobs or the need for our jobs, and it is these decisions that we have to make on a daily basis as a dispatcher that justify the need for so many of us, and its also whey we get paid the money that we do.
 
BruinsFan said:
The issue I have with your statement here is not that flight plans are inaccurate but that there are many factors that automation does not look at that a dispatcher would so even if a flight plan is accurate strictly from a 121 standpoint that doesn't mean it is a flight plan that I would sign my name to as a dispatcher. A great example is on transatlantic flights. Most automated flight planning systems don't look at Turbulence that closely and really don't take into account pilot reports or what the turbulence will do in the future, which are things almost any competent dispatcher will consider. So on a EWR-LHR flight for example, an automated system may plan a flight on NAT Track Z because that is the best wind route. However, a dispatcher will look at Track Z and see that there will be moderate turbulence between 40 and 20W. However if that flight were planned on NAT U, there would be much less turbulence and a better ride for the customers. I as a dispatcher would then plan it on NAT U. There was nothing inaccurate about the NAT Z flight plan, but NAT U was obviously a better choice.

And you don't take into account the stuff that happens after a flight is released. You could have ten flights from the northeast to Florida, and after they are checked over and sent ATC closes the AR Routes, and now all of a sudden you have to process reroutes for ten flights at once. This is why we are still needed. One other example I will give is I recently had a flight that couldnt land in ORD for windshear and my first instinct was to divert to STL. However, STL said they could not fuel the airplane, so I quickly had to change the diversion to DTW who could fuel the airplane. As far as the system goes, STL was legal. It's the things like this that lead me to believe that Automation will never eliminate our jobs or the need for our jobs, and it is these decisions that we have to make on a daily basis as a dispatcher that justify the need for so many of us, and its also whey we get paid the money that we do.
Click to expand...

In your examples it could be difficult to deal with some of these things with conventional programming, but that's not what I'm really talking about. You can actually handle the turbulence example conventionally now with an algorithm (find the most fuel efficient route with the least turbulence using existing turb API's), but I was talking about machine learning in the distant future. Machine learning collects historical data which builds a model that is able to adapt independently to new sets of data to produce reliable and repeatable results. So in your example it would be able to look at x years of turbulence data, flight outcomes, etc, and then look at the current data and produce results that in theory are the safest option. This is why training is an important part of ML, and any dispatch system built on ML technology would have to have an extensive initial learning phase. Yes you still would want somebody to review it, but if it has 90% or greater accuracy then there is very little to correct and every correction made helps to train the system for future flights. You would also still need a flight follower and the flight follower would need to make a decision on diversions. At the end of the day though an airline still needs significantly less people to do this job if such a systems exists.

Consider the work SpaceX did to make rockets land themselves. They used machine learning in order to create a convex optimization algorithm to help understand how to land the rocket, and real-time computer vision to identify the best routing. Yes, landing a rocket and dispatching a flight are entirely different but it speaks to the power of the technology which is only getting better.
 
You must log in or register to reply here.
Back
Top