Unknown Knowns
[Updated to fix some egregious writing errors.]
Recently, the WSJ's motoring correspondent, Dan Neil, made the case that we will cede control of our cars to the cloud sooner rather than later, and for the better. He starts by vividly demonstrating that a car can, in fact, pilot itself under demanding conditions:
In Mr. Neil's view, these autonomous driving systems (ADSs) "have long since passed the point of mere driving competence to arrive at something like expert status."
Therefore, it is only a matter of time before we hand over the controls, and the autonomy that implies, to these digital uberdrivers. After all, because ADSs will be able to predict rather than merely react, they would make the stop-and-go traffic jams a thing of the past. Highway carrying capacity will increase, because that same predictive ability will reduce following distances. Having pre-programmed the route, turbulence caused by lane changes will go away, and so will the need for almost all traffic control devices.
In contrast, human controlled vehicles will be outliers, the risk carriers. One of these days, we will be banned from driving on our own roads.
Despite being something of an gear head, I can see his point. But because I am a gear head, I find it disturbing that in some not too distant future, driving will be a lost skill. While that might seem an extreme conclusion, remember this: what was once taken for granted has already nearly vanished -- virtually no one under the age of 30 can drive a three pedaled car. That's a reduction of almost 100% within a couple generations.
Why? Because essentially no one among that group that makes almost all car buying decisions — women — enjoys telling the transmission what to do (full disclosure, both my cars are manuals). Similarly, hardly anyone views driving as any less a chore than vacuuming.
So we will, a few atavistic double-declutching stick rowers notwithstanding, happily cede the right of way to our digital overlords.
Left unanswered, though, is something of a conceptual problem. Granting that autonomously operated vehicles are the future, how do we get there from here?
There are two problems Mr. Neil doesn't address. First, AI should be re-branded as IDS: impenetrable digital stupidity. Despite decades of effort, as good as all things digital are at computation, they are completely sucktastic at learning.
That means ADSs are relying essentially on two things: vehicle positions and pre-programmed rules.
It is that last thing that is a bother. In driving, we know all sorts of "rules" — traffic rules, for instance. However, if asked how we drive, we couldn't fully articulate the process, because the are far more things we know about driving than we know we know (including how to juggle competing rules), and would even have a hard time articulating much of what we know we know.
Now, in the two dimensional driving space, perhaps that can be overcome to a sufficient degree. But then we get to another pothole: driver out-of-the-loop. This has been a perplexing problem as aircraft have become increasingly automated, to the point where (IMHO) most incidents and accidents are caused at least in part by pilot out-of-the-loop (e.g., AF447).
How can ADSs deal with unforeseen events outside their rule space, or when, for whatever reason, a car goes "off the reservation"? Obviously, by handing the controls back to the driver. Who may well be reading a book, or asleep, or …
So, I agree with Mr. Neil that ADSs will happen, but I doubt they will predominate in my lifetime, because these things only know what they have been told, which is less than what we need to tell them, and they are, and will remain, completely incapable of learning.
Some side notes:
I think fully APSs (autonomous piloting systems) are much further off. Operating in three dimensions is several orders of magnitude more difficult than driving, which means the problem of unknown knowns is worse. Further, going off the reservation will be both more likely, and consequential.
A couple personal examples. Very shorty after takeoff a couple years ago, the autothrust system pulled the power to idle without any warnings. Last year, on getting tight vectors to final, the flight management system turned the wrong way to intercept the instrument approach course. Absent in the loop pilots both those examples would have turned out very badly .
Another issue is complexity. The flight manuals for airplanes with flight management systems are roughly three times as heavy as for those who rely upon time honored, well worn, hand-tooled flying skills. That's one thing for pilots who are on the recieving end of extensive and expensive training. But what about those for whom having an iPhone in hand is to flirt with apoplexy (yes, TOSWIPIAW, I am looking at you)?
Recently, the WSJ's motoring correspondent, Dan Neil, made the case that we will cede control of our cars to the cloud sooner rather than later, and for the better. He starts by vividly demonstrating that a car can, in fact, pilot itself under demanding conditions:
The Mazda Raceway at Laguna Seca is a 2.2-mile asphalt roller coaster plunging and soaring across California's tawny Monterey highlands. The most famous section, the Corkscrew, requires drivers storming up a long hill to slam the brakes and take a hard left into what seems to be thin air. The car goes momentarily weightless, and when the track materializes beneath you—always a pleasant surprise—it's going downhill like a ski jump—and, oh yeah, heading hard right.
…
Except that I wasn't really driving. While I was indeed in the driver's seat, my hands and feet were weirdly unoccupied.
The car was driving itself, digitally duplicating a lap driven earlier by a professional driver—a man now sitting on the pit wall, watching the car and me come and go. All I had to do was sit there, with the car dancing on the edge of control under me, manfully freaking out.
BMW's TrackTrainer—an experimental 330i sedan bristling with machine-vision equipment—uses GPS, track maps and telemetry recorded during a professional driver's model lap to negotiate a racecourse.
In Mr. Neil's view, these autonomous driving systems (ADSs) "have long since passed the point of mere driving competence to arrive at something like expert status."
Therefore, it is only a matter of time before we hand over the controls, and the autonomy that implies, to these digital uberdrivers. After all, because ADSs will be able to predict rather than merely react, they would make the stop-and-go traffic jams a thing of the past. Highway carrying capacity will increase, because that same predictive ability will reduce following distances. Having pre-programmed the route, turbulence caused by lane changes will go away, and so will the need for almost all traffic control devices.
In contrast, human controlled vehicles will be outliers, the risk carriers. One of these days, we will be banned from driving on our own roads.
The danger will come not from auto-piloted vehicles but from the holdouts, those drivers who for whatever reason rely on the faulty, flimsy wetware between their ears. What will be normative? Should manually operated vehicles be the ones to give way? Or should autopilot cars (with special running lights) be especially deferential to their inferior human counterparts?
Despite being something of an gear head, I can see his point. But because I am a gear head, I find it disturbing that in some not too distant future, driving will be a lost skill. While that might seem an extreme conclusion, remember this: what was once taken for granted has already nearly vanished -- virtually no one under the age of 30 can drive a three pedaled car. That's a reduction of almost 100% within a couple generations.
Why? Because essentially no one among that group that makes almost all car buying decisions — women — enjoys telling the transmission what to do (full disclosure, both my cars are manuals). Similarly, hardly anyone views driving as any less a chore than vacuuming.
So we will, a few atavistic double-declutching stick rowers notwithstanding, happily cede the right of way to our digital overlords.
Left unanswered, though, is something of a conceptual problem. Granting that autonomously operated vehicles are the future, how do we get there from here?
There are two problems Mr. Neil doesn't address. First, AI should be re-branded as IDS: impenetrable digital stupidity. Despite decades of effort, as good as all things digital are at computation, they are completely sucktastic at learning.
That means ADSs are relying essentially on two things: vehicle positions and pre-programmed rules.
It is that last thing that is a bother. In driving, we know all sorts of "rules" — traffic rules, for instance. However, if asked how we drive, we couldn't fully articulate the process, because the are far more things we know about driving than we know we know (including how to juggle competing rules), and would even have a hard time articulating much of what we know we know.
Now, in the two dimensional driving space, perhaps that can be overcome to a sufficient degree. But then we get to another pothole: driver out-of-the-loop. This has been a perplexing problem as aircraft have become increasingly automated, to the point where (IMHO) most incidents and accidents are caused at least in part by pilot out-of-the-loop (e.g., AF447).
How can ADSs deal with unforeseen events outside their rule space, or when, for whatever reason, a car goes "off the reservation"? Obviously, by handing the controls back to the driver. Who may well be reading a book, or asleep, or …
So, I agree with Mr. Neil that ADSs will happen, but I doubt they will predominate in my lifetime, because these things only know what they have been told, which is less than what we need to tell them, and they are, and will remain, completely incapable of learning.
Some side notes:
I think fully APSs (autonomous piloting systems) are much further off. Operating in three dimensions is several orders of magnitude more difficult than driving, which means the problem of unknown knowns is worse. Further, going off the reservation will be both more likely, and consequential.
A couple personal examples. Very shorty after takeoff a couple years ago, the autothrust system pulled the power to idle without any warnings. Last year, on getting tight vectors to final, the flight management system turned the wrong way to intercept the instrument approach course. Absent in the loop pilots both those examples would have turned out very badly .
Another issue is complexity. The flight manuals for airplanes with flight management systems are roughly three times as heavy as for those who rely upon time honored, well worn, hand-tooled flying skills. That's one thing for pilots who are on the recieving end of extensive and expensive training. But what about those for whom having an iPhone in hand is to flirt with apoplexy (yes, TOSWIPIAW, I am looking at you)?
posted by Hey Skipper at 12:05 PM
18 Comments:
"How can ADSs deal with unforeseen events outside their rule space?"
My 16 year old daughter recently got her license and all I can say is, "Holy _____!!!!" Need I say more?
The typical driver is not like a professional pilot. They don't have years of training, a gazillion hours of practice, and multiple certifications. Nor do they focus on driving with the seriousness and responsibility that a professional pilot has when he flies a plane with hundreds of people in it.
Nor can a car be slammed out of the sky by downdrafts associated with thunderstorms. Nor can an airplane just stop in the middle of the air like a car can on the ground. Nor does a car need to deal with that 3rd dimension. Nor is a car going 600 mph. Etc.
It also seems to me that outliers in a large jet have significant delays built in (though I'm not sure of this). For example, when I hit the brake, my car nearly instantaneously starts slowing while it seems to me a large plane is more like a boat in that when you ease off the throttle there's more of a delay before anything starts happening.
In the case of outliers in the car, it seems to me that you either react nearly instantaneously or forget it, the accident has already happened so there won't be switching back to the driver if the car computer doesn't know what to do.
But google already has millions of miles on their cars and while computers may not be good at learning outside the box (though neither are poeple), they're fabulous at matching current conditions or events to a massive database of past conditions/events and quickly recalling what worked those times.
Will there be events that are outside both the rules and database? Sure, but then the car will simply slow down as quickly as it can until either normalcy returns or it crashes at hopefully a much slower speed. Once most cars are autonomous, they can communicate what they're doing to the surrounding vehicles so they have a heads-up and can avoid the confused car.
Again, the computer is competing against human drivers who are texting somebody while looking at the map and talking to someone else on the phone and chatting with the passenger next to them driving with their left leg with a beer between their thighs after only getting 4 hours of sleep the night before.
My bet is that by 2020 the computer can do consistently better and by 2030 it will be cheap.
I can often tell when a driver is going to run a red light by the tilt of his head. I bet autonomous cars can't do that
This comment has been removed by the author.
As an aside - manual cars are still the norm outside the USA. I think you can look to the rest of the world to maintain those skills for a good while yet.
Heh. Imagine a hundred years from now when the big deal for historical recreators is to learn to drive a manual transmission.
Bret:
You are correct in all your suppositions about what makes autonomous flying much harder than driving.
But I'm not sure some of the other stuff has been fully taken on board. For instance, within a fairly narrow temperature range, a road can go from being merely wet, to glare ice (and the air temp doesn't necessarily have to be below freezing). How does an autonomously driven car recognize the difference?
One way is how some people do it -- ex post facto, while in some awkward location in the midst of some unfortunate landscape. However, most people can see the difference, and compensate.
Of course, because it is something humans can sense, I'm sure we can design something to do the same thing. But at what cost? Will the sensors have to be on all roads with any possibility of icing, or on each car?
Harry:
The fascinating part about this (and which serves as part answer to my questions) is that, to some extent, while the net can't learn, it can know: it would know well before you whether the prospective red-light running car was slowing as it should.
Which leads to a prediction: the first implementation of APS will be cars knowing the location of traffic control devices, and what the devices' state will be when the car arrives. Combined with knowing the range and relative speed of the car in front, accidentally running stop signs red lights, and turning in front of oncoming traffic, will progressively become a thing of the past.
In fact, some high-end cars now have adaptive cruise control that will even bring the car to a stop, if need be. (It is the sort of thing, BTW, that is a perfect example of trickle down economics, but which the left resolutely ignores, despite there being thousands more like it).
Mark:
Now that you mention it, a few years ago, I rented a car in England. It was a manual, as were probably most on the lot.
I'll bet there isn't one non-specialty rental car in the entire US that has a manual.
Why is that?
I don't know about that adaptive cruise control trickling down. We bought a minivan in 2003 and had that feature. It was wonderful. But SWIPIAW hit a deer and we had to replace the minivan. That feature was available from only one manufacturer and then only with the absolutely maxed out package set. So now we don't have it. We asked the car guy about it and he said it just wasn't popular, nobody (except freaks like us) wanted it.
Trying to predict the popular appeal of cool emerging technologies is usually a mug's game, somewhat akin to predicting climate or resource stocks decades into the future. Voice activated dictation for word-processing has been around for a long time, forever "improving". Every few years my associates invest in new software for it, marvel at the time and money it saves them, recommend it to one and all and then abandon it after a couple of months. I could be out of the loop, but I'm not under the impression Kindle is sweeping the nation.
OTOH, there was that IBM VP who foresaw little general interest in personal computers and I doubt even Noika predicted the ubiquity of the cellphone.
This one may prove to be a top-down imposition by traffic planners on a sceptical and resistant public. The spectre of abolishing traffic jams in one fell swoop must be irresistible, but I can't see anyone over thirty being able to relax with this and trust it. Can the driver sleep or drink with it? Will it be illegal to turn it off and slow traffic down? Will employers now expect us to make productive time of our commutes? The world may be changing fast, but surely it will take a very long time before everyone has a common percepton of how safe it is. Bret's list of reasons why automatic pilots in aircraft still require close and alert human supervision is all true, but one thing he forgot to mention is that when automatic pilot systems fail, other aircraft aren't usually whizzing by one after the other six feet away.
Peter wrote: "...when automatic pilot systems fail, other aircraft aren't usually whizzing by one after the other six feet away."
No, but the ground (sometimes in the form of mountains) is often whizzing away at hundreds of miles an hour. After all, landing an aircraft is really just a controlled collision of the aircraft with a planet, though I doubt Skipper views it quite that way. :-)
The "one fell swoop" thing is likely incorrect. The initial demand will be overwhelming in certain areas. The elderly when they can really no longer drive themselves safely, taxicabs and commercial, long-haul trucking are some examples of where autonomous vehicles might first be used. Once a toehold is attained in any area and safety is adequately proven, it will incrementally, over decades, spread to the rest of vehicles.
In the '50s, we were all going to have personal helicopters, with no thought about traffic control.
I wonder about autonomous parking in cramped areas.
In the '50s, we were all going to have personal helicopters, with no thought about traffic control.
I wonder about autonomous parking in cramped areas.
Parking in tight quarters is the scenario where autonomic driving systems are most likely to outperform humans.
Peter:
You are right, but I don't think there needs to be anything like a "one fell swoop" implementation.
For instance, it is possible for GPS nav systems to know where all stop signs are, and to first warn, then take over if the driver isn't slowing in time. Similarly, it is easy to envision traffic signals communicating their current and future states to those same nav systems, which could then calculate the most efficient speed, while providing the same warnings as for stop signs.
Neither of these would be expensive in quantity, or would force drivers to cede control (except during lapses in situational awareness). Nor would they exclude legacy vehicles. The incentive would be the break on insurance costs -- probably the single greatest contributor to the accident rate is intersection collisions.
Bret's list of reasons why automatic pilots in aircraft still require close and alert human supervision is all true, but one thing he forgot to mention is that when automatic pilot systems fail, other aircraft aren't usually whizzing by one after the other six feet away.
That's true. However, there is a single overwhelming difference. The analogy is how chickens and pigs view breakfast: one is involved, the other committed.
In a car, if you get lost, or get a flat, or the weather gets bad, or the ADS has a fit, you can simply pull to the side of the road and stop. In contrast, the moment you takeoff, you are committed to shedding the last bits of energy at very specific places; otherwise, things will go very, very badly.
Bret:
Once a toehold is attained in any area and safety is adequately proven, it will incrementally, over decades, spread to the rest of vehicles.
As I mentioned above, high end cars have adaptive cruise control (googling around a bit, I found a BMW discussion board where people were raving about how good it is), lane departure warning and assisted braking.
All of that will be common in a few years, and none of it will be noticed by the left, whose endless ridicule of trickle down is focused on a single, dimensionless number (and even that one is wrong).
"high end cars have adaptive cruise control" - again, it was available on middle range cars 10 years ago, and now isn't. We had to do without because the level of car we buy is no longer high end enough for it. So the adoption, if any, is rather slow.
Eight manufacturers, all high end, provide full authority ACCs.
Four more have plans to do so in the next two years, two of them, Subaru and Chevy are mid range.
Eleven more sell limited authority ACCs (which is probably what you had, pre-deer), six of which are mid-range.
(source)
Skipper:
I don't doubt that you, Bret and the WSJ author, who clearly can hardly wait for all the stodgy old resisitng wrinklies to pass on, will ultimately prove right, especially if the personal financial inducements are as significant as you forsee. But I wonder whether it may take longer than you think, not because of the technology, but because this seems to be challenging an element of timeless human nature.
Scientists and experts tend to assume that, once their studies and statistics "prove" that something is safer than something else, human behaviour will adjust with reasonable speed. It sometimes does, but not always. Endless studies have shown that flying is statistically safer than driving, but nervous fliers are more common than nervous car passengers. We all know that our chances of being killed driving to work are greater than being killed in a terrorist attack. Yet although we make pragmatic trade-offs with driving, we invest billions and accept serious freedom/civil liberties restrictions to prevent even one person being killed in an attack. Endless studies have shown that women and children are at greater risk of a sexual assault within their family circles, but it's the spectre of a sicko stranger hiding in park bushes that informs much of our thinking and our nightmares.
My theory is that these "anomolies" can be explained with reference, not to doubting the experts, but to fear of the unknown and a loss of control, or even the illusion of such a loss. My priority, frankly, is not national accident statistics, it's my and my family's safety. I believe, perhaps naively, that they are safer because of my driving skill and ability to react in a crisis. It simply wouldn't be enough to tell me the car will stop or even turn when in danger, or that I am statistically 17% safer than I used to be. I would have to be assured, both cerebrally and viscerally, that it can instantly swerve out of the path of a runaway truck barrelling towards me without driving into a ditch or off a cliff. And if there are too many publicized tragic failures during the development stage, I'm not going to relax just because some boffin assures me they've fixed the bugs.
I once was flying to Northen Quebec on a 737 on one of those humid, turbulent days of late summer. Cloud were dark, moving fast and so thick we were still bouncing in and out of them at cruising altitude. The pilot was trying to adjust and there were lots of sudden speed changes and strange noises from the engines. The guy beside me was extremely nervous and agitated, seized up repeatedly and muttered things like "God, I hate it when he does that." It was not pleasant.
After an hour we were clear and enjoying dinner. We struck up a conversation and I asked him what he did. He said we was a pilot for the airline going north for a flight the next day. Needless to say, I was grateful not to have known that an hour earlier.
So call me when the truckers of America endorse this unanimously.
Peter wrote: "But I wonder whether it may take longer than you think, not because of the technology..."
I've been giving robotics talks since about 2000 where I've predicted that the technology for autonomous vehicles would be pretty good by 2020. In all of those talks I pointed out that other factors like liability and societal comfort could delay the widespread adoption of such vehicles by many decades,
So I agree with you.
[Peter:] I don't doubt that you, Bret and the WSJ author, who clearly can hardly wait for all the stodgy old resistng wrinklies to pass on, will ultimately prove right …
I must have lost track of my point somewhere after the title. I don't Mr. Neil is right, for two reasons. 1. The act of driving is essentially known; drivers make mistakes, but they aren't due to something that wasn't knowable at the time. But we don't know everything we know, and much of what we do know we know isn't readily explainable. 2. There is practically no such thing as machine learning, and at the rate of progress in that realm, there may never be. Since rote behavior and driving don't mix, a full authority ADS will be a non-starter.
I can't help but notice that the ADS Mr. Neil was praising was a perfect example of rote behavior. After all, it isn't like the thing figured out Laguna Seca on its own, and it wouldn't have the first idea of what to do if a car in front of it blew an engine — hey, it's a race track, these things happen — and laid some coolant on the track. Nor, after having ended up somewhere in the landscape, would it have any new ideas on how not to get there again.
Post a Comment
<< Home