Go Back  PPRuNe Forums > Flight Deck Forums > Tech Log
Reload this Page >

Computers need to know what they are doing

Tech Log The very best in practical technical discussion on the web

Computers need to know what they are doing

Old 17th Aug 2016, 08:41
  #41 (permalink)  
 
Join Date: Mar 2006
Location: USA
Posts: 2,509
Likes: 0
Received 0 Likes on 0 Posts
If I were guessing I would guess that you people already have this technology and the government and the airlines have made you swear an oath that you will keep it secret
We're not allowed to discuss that sort of sensitive security information on a public message board.
Check Airman is online now  
Old 17th Aug 2016, 09:08
  #42 (permalink)  
 
Join Date: Jan 2013
Location: Australia
Age: 46
Posts: 83
Likes: 0
Received 0 Likes on 0 Posts
Yes, it would be fantastic if aircraft could "run a simulation" and foresee all possible outcomes of everything that could happen to them in the next minute, and automatically react accordingly.

It would also be fantastic if airliners were doing Mach 3 everywhere at half the current fuel burn. Why don't we just do that now?
BleedingAir is offline  
Old 17th Aug 2016, 09:35
  #43 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
I didnt say all possible outcomes of everything that could happen to them

anything can be ridiculed if you exaggerate it to the point of absurdity

how many possibilities are there in a chess game?
yet computers play chess just fine

Last edited by em3ry; 17th Aug 2016 at 09:50.
em3ry is offline  
Old 17th Aug 2016, 12:09
  #44 (permalink)  
 
Join Date: Oct 2012
Location: Australia
Posts: 85
Likes: 0
Received 0 Likes on 0 Posts
Chess is a vastly simpler proposition though. The moves the player can make are limited and discrete. The moves the opponent can make are also limited and discrete. At the start of a chess game, for example, there are twenty available moves for White (eight pawns can each move one or two spaces, two knights can each go to two positions). For each of these moves the opponent can also make twenty moves, so to look ahead one turn requires 400 separate evaluations. The second turn is a lot more complex; depending on which pawns moved you might be able to shift any of the other pieces. Depending on how the opponent moved there might even be a capture. Still, we're talking about a few billion moves to be evaluated after a few moves, and with a bit of alpha-beta pruning (practical when "score" is reasonably well defined, for example by assigning values to pieces) this is quite manageable on a modern PC.

Running a simulation on a transport aircraft using fixed control inputs and in perfectly known conditions is practical. Obviously you'd have to make some sacrifices in aerodynamic modelling to have it done by an aviation-grade computer (not exactly the latest and greatest desktop hardware) much, much faster than realtime (there's not much point if at any time you can only look a few seconds ahead) but I can see that being possible.

Then you start throwing in variables. What happens if we apply 5% more thrust? What happens if we apply 10% more? Or 15%? Or maybe 14.3726% on the left engine but 1.3978% on the right since that might be a practical way to deal with a rudder malfunction? Or should we apply 5% left aileron? Or 5% right rudder? Or drop the gear?

What if there's a bit of wind shear? Maybe some wake turbulence? Maybe the C-17 that's about to taxi onto our runway will realise what's going on, apply reverse thrust, and get out of the way. Or maybe he'll stop, so that just staying a bit to one side provides sufficient safety margin. Or maybe he'll keep on coming, in which case absolutely any response that involves continuing this landing is going to cause a disaster. Of course, for the first option the computer has to be able to recognise that (a) the plane is a C-17, and (b) C-17s can to taxi backwards on reverse thrust. And if it's listening on the radio it might be able to hear ATC telling the C-17 to stop, and the C-17 pilot confirming (or not).

So now instead of running one simulation to see what will happen, the computer is running thousands or millions of simulations just for the next millisecond. And for each of those simulations, thousands or millions for the next millisecond (so now we're looking at millions to trillions). By the time you're looking maybe 30 seconds ahead (which is probably about the minimum you could expect) you're asking for a *lot* of simulations - and of course because these simulations are covering the dynamics of a large aircraft each one is pretty complex!

Plus, of course, you need big sanity-checks. If someone's gone and parked a plane on the runway you were about to land on, then the best strategy for the next 30 seconds may well be to cut power, lower flaps, lower gear, and raise the nose. All of these will reduce the speed, and therefore prevent a crash within the 30-second simulation time. What the simulation does not know (because it only looks 30 seconds ahead) is that at 31 seconds the plane will stall and crash, and the burning wreckage will slide into the fully-fuelled plane sitting on the runway. You can force the computer to only accept "nice" solutions (eg. where the plane finishes well above stall speed, well below VNE, within 30 degrees of level in both axes, etc) but then that may cut out beneficial solutions that a human pilot could find (eg. coming very close to a stall in order to pull up fast enough). The other effect of the 30-second look-ahead time is that the computer won't even start to think about the guy sitting on the runway until it's 30 seconds in the future; a human might have seen that sveral minutes ago.



The first part of this plan, controlling the plane itself, is essentially heading for a field known as "optimal control" - how do you perform control in the absolute best way to achieve the goal. One way, as you've described, is to brute-force every possible action and pick the best. The downside, of course, is that this tends to take many times the age of the universe on any conventional computer that can be constructed within the universe. The fact that there's a field of study rather than a single "optimal control equation" indicates that nobody's actually solved optimal control for any but the most trivial cases. We do have numerical methods that can produce a reasonable approximation for more complex systems, but a plane is a very complex system which will need a very large amount of careful analysis.

Add to that the outside factors. What are other planes doing? What is the air around the plane doing? Can ATC deal with your crazy plan? What happens if your plane suffers a mechanical failure and so the best plan you had no longer works?

If you can assign numbers to the probability and impact of every possible event, and then spend a long time crunching numbers, you can undoubtedly come up with a single best course of action. By the time this has occurred it's likely that the human race will have become extinct - and that's a very big "if" on assigning the numbers.


Or you can just go with the existing solution, where if the problem is one that the computer knows about then it just implements a solution, and otherwise it hands over to a human. The computer already knows about most of the reasonably common problems, and handles them so well that nobody even notices. Humans are good at spotting future problems well in advance by collating extra information; if there's someone entering the runway that you're about to land on, you'll be recalling what they and ATC have been saying to determine whether it's a risk or not.

Of course, there's still a class of problems that can't be seen coming but do require an instant response, like the 737 rudder PCU malfunctions. We don't have a good solution to this, but I'm not sure that a simulation would help at all. After all, how can you simulate a malfunction when you don't know what the malfunction is?
Slatye is offline  
Old 17th Aug 2016, 12:58
  #45 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
Flying a plane should be simpler than playing chess. You know the exact route that you want to take. You just don't know the exact moves that will take you along that exact route
em3ry is offline  
Old 17th Aug 2016, 14:32
  #46 (permalink)  
 
Join Date: Jun 2006
Location: Brisbane
Posts: 265
Likes: 0
Received 0 Likes on 0 Posts
Em3ry,

Your posts make it clear that you have no knowledge of piloting an aircraft. Some of us who are pilots also know a little about computers. I have studied university level AI. You obviously haven't. You've read an article about the Google car and have come onto a professional pilot's forum telling everyone how to fix the industry. You are way out of your depth.

One speculated cause of the latest crash was a man-machine interface issue. I'm not aware of a similar accident in recent times, so it's a rare one.

As I have said, this will continue to be an issue, but rushing automation will not be the solution. In fact, it is very likely the cause.

Yes, more advanced warning systems and improved man-machine philosophy and training will continue to improve safety, but in aviation it is a slow process because a new technology must be well proven and certified before it can be implemented (a bit like medical science). If it takes an automotive company 5 years to develop and certify a lane-warning system on a car, you can bet it would take 15 years to do something similar on an aircraft.

The state of the industry at present (and for the forseeable future) is that it is very difficult for a human to hand fly an aircraft all day safely. It is also impossible for a computer to do the same. So the best solution is a combination of the two. It is the man-machine interface that we are working on, and it's not easy. Financial pressure on training, pay, conditions, fatigue etc also influence the human element, often leading to suggestions that increased automation is the answer. Widely held recent opinion is that it is not.

Last edited by Derfred; 17th Aug 2016 at 15:04.
Derfred is offline  
Old 17th Aug 2016, 15:04
  #47 (permalink)  
 
Join Date: Dec 2001
Location: what U.S. calls ´old Europe´
Posts: 941
Likes: 0
Received 0 Likes on 0 Posts
Computers, as we know them today are great at doing repetitive tasks and monitoring. Decision-making, no so much.
An interesting read...
IBM's Watson is better at diagnosing cancer than human doctors
I would call this (especially when it comes to decide about the type of treatment) some sort of decission making.
However, this decission is purely statistical and based on comparison with millions of datasets. So our first step is that we need to install very accurate flight data recorders, fly for some centuries and after having a database of millions of accidents, the computer will be perfect in predicting an impending crash from the current parameters.
Maybe it will even be able to decide upon an action, which in the past had saved the day. However, we need to freeze airports, airspace and aircraft design for some time to collect relevant data...
Volume is offline  
Old 17th Aug 2016, 15:18
  #48 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
If it takes an automotive company 5 years to develop and certify a lane-warning system on a car, you can bet it would take 15 years to do something similar on an aircraft.
That's why they should get to work on it right away
em3ry is offline  
Old 17th Aug 2016, 16:34
  #49 (permalink)  
BBK
 
Join Date: Feb 2001
Location: UK
Posts: 468
Received 1 Like on 1 Post
em3ry

If you want to advocate that a computer system can do the job of a pilot then you need to understand what the job entails. You appear not to. Or maybe you do and you're a troll and trying to wind us up.

So much of what flight crews do involves a balanced judgement rather than black and white decisions. Disruptive passengers, in flight medical emergencies, weather avoidance etc.
Of course this is all goes on behind a locked cockpit door unseen to the passengers but that's no excuse for not having the imagination to think about such matters if you want to argue that computers can do a better job.

I read an article recently by a pilot on this subject and he made a very good comparison with the medical profession in that technology has improved medical care but we never think that doctors will be replaced by robots.

BBK
BBK is offline  
Old 17th Aug 2016, 17:20
  #50 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
all I said was computers can be made smarter and that doing so would have prevented some of the recent accidents

Last edited by em3ry; 17th Aug 2016 at 17:42.
em3ry is offline  
Old 17th Aug 2016, 17:42
  #51 (permalink)  
BBK
 
Join Date: Feb 2001
Location: UK
Posts: 468
Received 1 Like on 1 Post
em3ry

I was going to reply but Derfred's posting above pretty much sums up my views. Computers, in a well designed system, can enhance flight safety but not, IMHO, in the role of decision making.

Just my two penn'orth.

BBK
BBK is offline  
Old 17th Aug 2016, 17:46
  #52 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
which is why you need a pilot
em3ry is offline  
Old 17th Aug 2016, 17:50
  #53 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
as i said earlier
Originally Posted by em3ry
The human pilot will always be better than the computer at certain tasks.
But the computer will always be better than the human pilot at certain other tasks
em3ry is offline  
Old 17th Aug 2016, 17:51
  #54 (permalink)  
BBK
 
Join Date: Feb 2001
Location: UK
Posts: 468
Received 1 Like on 1 Post
Pilots. Plural! Over and out.
BBK is offline  
Old 17th Aug 2016, 18:25
  #55 (permalink)  
 
Join Date: Jul 2013
Location: Everett, WA
Age: 68
Posts: 4,356
Received 157 Likes on 75 Posts
The FAA is on record has prohibiting the use of anything resembling AI for aircraft or aircraft engine control. That's right, a flat prohibition. The reason is that AI can be somewhat unpredictable - and that's the very last thing you want to integrate into an aircraft system.
Personal example - my BMW was a very basic form of AI that 'learns' my driving style and adapts the engine and trani controls accordingly. A while back, they needed to re-flash the BMW computer to correct a malfunction. When I picked the car up, it had turned into a gutless wonder - literally unsafe - I nearly got in an accident on the way home when it failed to accelerate when I expected to - so I took it back next day. Turns out the AI values had gotten corrupted when they did the re-flash. They reset the AI values and the car was fine. A co-worker had a similar experience with his car (non-BMW).
It's going to be a long time before AI has advanced to the level where we're going to be using to control aircraft.
tdracer is online now  
Old 17th Aug 2016, 18:28
  #56 (permalink)  
 
Join Date: Mar 2006
Location: USA
Posts: 2,509
Likes: 0
Received 0 Likes on 0 Posts
Originally Posted by em3ry
Flying a plane should be simpler than playing chess. You know the exact route that you want to take. You just don't know the exact moves that will take you along that exact route
And with that, I'm leaving this discussion. Clearly a troll.
Check Airman is online now  
Old 17th Aug 2016, 18:40
  #57 (permalink)  
Thread Starter
 
Join Date: Sep 2013
Location: California
Posts: 37
Likes: 0
Received 0 Likes on 0 Posts
Is the FAA on record as forbidding running a simulation in real time with current events to see what is just about to happen if only to warn the pilot
em3ry is offline  
Old 17th Aug 2016, 19:30
  #58 (permalink)  
 
Join Date: Feb 2016
Location: Southport
Posts: 1,333
Likes: 0
Received 7 Likes on 6 Posts
How do you ensure that your "simultaneous simulation" is capable of predicting the future, as that's effectively what you need it to do?
andytug is offline  
Old 17th Aug 2016, 20:09
  #59 (permalink)  
 
Join Date: May 2000
Location: Seattle
Posts: 3,195
Likes: 0
Received 0 Likes on 0 Posts
And all of this is relevant to my original post. All I'm suggesting is a way of making the computers smarter.
Are you people seriously suggesting that smarter computers would be a bad thing?
. . .

Is the FAA on record as forbidding running a simulation in real time with current events to see what is just about to happen if only to warn the pilot
You are not suggesting anything new. Computers are constantly improved to make them "smarter".

As has been pointed out before, even computers in airplanes have been improved to make them smarter. They have even been programmed with the capability to assess the current situation, predict possible outcomes, AND warn the pilot - all with the FAA's blessing.

So, unless you have a SPECIFIC proposal to perform a SPECIFIC prediction to generate a SPECIFIC set of responses or warnings, you are merely repeating a baseless proposition ad nauseum. So far, it is apparent you do not have that capability...
Intruder is offline  
Old 17th Aug 2016, 23:26
  #60 (permalink)  
 
Join Date: Aug 2007
Location: Utterly insignificant little blue-green planet, unregarded yellow sun, unfashionable end, western spiral arm, Milky Way
Age: 38
Posts: 276
Received 0 Likes on 0 Posts
Originally Posted by Tourist
10 If height => 20000ft then ignore EGPWS.

Problem solved.

Next?
Ah, so, bang into Mt. Denali and the Himalayas, then. Brilliant...
semmern is offline  

Thread Tools
Search this Thread

Contact Us - Archive - Advertising - Cookie Policy - Privacy Statement - Terms of Service

Copyright © 2024 MH Sub I, LLC dba Internet Brands. All rights reserved. Use of this site indicates your consent to the Terms of Use.