Re: [LINK] Cars, again

[Tom Worthington]

On 07/12/17 10:27, Jim Birch wrote:


XKCD on self-driving cars ...


Ken Kroeger, Executive Chairman of Seeing Machines gave a fascinating 
presentation on Monday night in Canberra about their technology for 
self-driving cars. Seeing machines makes head and eye tracking 
technology, originally developed at ANU, to check drivers of vehicles 
are paying sufficient attention: 
http://blog.tomw.net.au/2017/12/canberra-start-up-overnight-success.html


While Ken was speaking in Canberra, his colleagues in London were 
raising 35 million Pounds for expansion: 
http://www.iii.co.uk/alliance-news/1512494249463655700-3/seeing-machines-raises-gbp35-million-in-oversubscribed-placing-alliss-


This technology could make ordinary vehicles safer by checking if the 
driver is paying sufficient attention to the road. If they are looking 
down at a smart phone, or at the kids in the back, a warning tone could 
be sounded from the direction the driver should be looking.



--
Tom Worthington, MEd, FHEA, FACS CP. http://www.tomw.net.au t 
+61(0)419496150

TomW Communications Pty Ltd. PO Box 13, Belconnen ACT 2617, Australia
Liability limited by a scheme approved under Prof. Standards Legislation

Honorary Senior Lecturer, Computer Science, Australian National University
https://cecs.anu.edu.au/research/profile/tom-worthington
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Jim Birch]

XKCD on self-driving cars:

https://xkcd.com/1925/


On 17 November 2017 at 21:19, David Boxall  wrote:

> On 17/11/2017 9:57 AM, David Lochrin wrote:
>
>> On Fri, 17 Nov 2017 07:16:08 David Boxall wrote:
>>
>>> Nobody really knows what we're doing, but we _are_ doing it. Some will
>>> huddle in a corner, whimpering. Others will get on with the job.
>>>
>> And some will happily sit back and watch the circus unfold! ...
>>
> Meanwhile, back in the real world: an electric semi with a range of 800 km
> and 30-minute recharge time that outperforms a diesel, plus a 400 km/h
> electric roadster with a 1,000 km range. Both have Tesla's driver
> assistance system, which is heading toward autonomy.
> http://reneweconomy.com.au/elon-musk-unveils-the-long-range-
> tesla-semi-electric-truck-19790/
>
> Fear might keep us safe, but it doesn't do a lot to advance humanity.
>
> --
> David Boxall|  The more that wise people learn
> |  The more they come to appreciate
> http://david.boxall.id.au   |  How much they don't know.
> --Confucius
>
> ___
> Link mailing list
> Link@mailman.anu.edu.au
> http://mailman.anu.edu.au/mailman/listinfo/link
>
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David Boxall]

On 17/11/2017 9:57 AM, David Lochrin wrote:

On Fri, 17 Nov 2017 07:16:08 David Boxall wrote:
Nobody really knows what we're doing, but we _are_ doing it. Some 
will huddle in a corner, whimpering. Others will get on with the job. 

And some will happily sit back and watch the circus unfold! ...
Meanwhile, back in the real world: an electric semi with a range of 800 
km and 30-minute recharge time that outperforms a diesel, plus a 400 
km/h electric roadster with a 1,000 km range. Both have Tesla's driver 
assistance system, which is heading toward autonomy. 
http://reneweconomy.com.au/elon-musk-unveils-the-long-range-tesla-semi-electric-truck-19790/


Fear might keep us safe, but it doesn't do a lot to advance humanity.

--
David Boxall    |  The more that wise people learn
                    |  The more they come to appreciate
http://david.boxall.id.au   |  How much they don't know.
--Confucius
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Fri, 17 Nov 2017 07:16:08 David Boxall wrote:

> Nobody really knows what we're doing, but we _are_ doing it.
> Some will huddle in a corner, whimpering. Others will get on with the job.

And some will happily sit back and watch the circus unfold!

David L.

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David Boxall]

On 16/11/2017 11:02 PM, Karl Auer wrote:

...

The practical question of implementation only becomes relevant when
there's agreement about the goal.


No new technology has ever - EVER - followed that path. There will be
no agreement about goals, no agreement about anything. People will get
on with implementing stuff, using thousands of different "benchmarks",
while the law and regulators follow years behind.
...

From Musk himself:

“Regulators may require some significant margin above human capability in order 
for a full autonomy to be engaged,” said Musk. “They may say it needs to be 50 
percent safer, 100 percent safer, 1000 percent safer, I don’t know. I’m not 
sure they know either.”

...

“Now that the foundation of the Tesla vision neural net is right, which was an 
exceptionally difficult problem, as it must fit into far less computing power 
than is typically used, we expect a rapid rollout of additional functionality 
over the next several months, and are progressing rapidly towards our goal of a 
coast-to-coast drive with no one touching the 
controls.”https://www.inverse.com/article/38049-elon-musk-self-driving-autopilot-tesla


If we truly want to examine accidents involving autonomous vehicles, 
then there's far more to learn from this:

Las Vegas police officer Aden Ocampo-Gomez said the truck’s driver was at fault 
for the crash ...

https://www.theguardian.com/technology/2017/nov/09/self-driving-bus-crashes-two-hours-after-las-vegas-launch-truck-autonomous-vehicle
The autonomous shuttle could have avoided the accident, but wasn't 
programmed to back away from the hazard. In the lab staying still 
probably seemed safer, but it assumed the human driver was paying attention.


Nobody really knows what we're doing, but we _are_ doing it. Some will 
huddle in a corner, whimpering. Others will get on with the job.


--
David Boxall|  For when the One Great Scorer comes
|  To mark against your name,
http://david.boxall.id.au   |  He writes-not that you won or lost-
|  But how you played the game.
  --Grantland Rice
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Karl Auer]

On Thu, 2017-11-16 at 22:19 +1100, David wrote:
> That won't work because you're asking people

No I'm not. They will do it themselves, in droves, as soon as anything
halfway useful comes on the matrket. As they are already doing with
Tesla.

>  to place the lives of themselves and their families in an opaque
> piece of technology with the vague assurance that, on average, it
> will be good for the accident statistics.

Theywon't do it because of assurances. They will do it because they
want the features. And as for putting their lives in opaque pieces of
technology, everyone does that every day.

> The practical question of implementation only becomes relevant when
> there's agreement about the goal.

No new technology has ever - EVER - followed that path. There will be
no agreement about goals, no agreement about anything. People will get
on with implementing stuff, using thousands of different "benchmarks",
while the law and regulators follow years behind.

>  And the goal needs to be expressed in ethical terms because the
> technology is so opaque and human lives are at stake.

You can express your goals any way you like, but if you can't measure
how close you are to achieving them, it's waffle.

> (1)  The vehicle is to be able to climb a (given) mountainous road
> with no safety fence to mark the edge on a rainy night at a speed of
> at least 'v' kph, and must do so successfully on 100 consecutive
> attempts.  That might test the vehicle's sensors.

So - it has to be WAY better than most human drivers. Uhuh. On a
mountain road atypical of most mountain roads. In particularly
difficult circumstances. That a competent human driver would avoid.

> (2)  The vehicle is to come to a complete stop from a speed of 'x'
> kph in 'y' seconds when an obstruction is suddenly placed 'z' metres
> in its path, and must do so on 100 consecutive attempts.

Again, way better than most human drivers. Though this task is probably
one where the average automaton could be way better than most human
drivers.

> ...that sort of thing.  What else would you expect an expert driver
> to be able to do?

I want a competent, dependable, reliable, law-abiding driver who
concentrates on the task at hand. One who says "why don't we go up that
mountain tomorrow in daylight when the rain has stopped?"

>   We need another use-case to test the vehicle in an environment
> where there are many vehicles doing erratic things at high speed, for
> example in a multiple vehicle accident where a truck suddenly diverts
> into an oncoming line of traffic.  I'm thinking of the accident on
> the Hume Highway a little time ago when a truck did just that.

And how did the humans do?

You want perfection, or at least extreme superiority. But autonomous
vehicles only need to be just a bit [safer, faster, cheaper] than human
drivers to make the world a better place. The "just a bit" will become
"quite a bit" and eventually there will be no rational reason for
humans to drive vehicles any more, except for fun, in isolated
environments with like-minded enthusiasts.

It all starts with "just a bit". And there is no way to measure how big
that bit is, except via a statistical examination of the facts - after
the fact.

Regards, K.


-- 
~~~
Karl Auer (ka...@biplane.com.au)
http://www.biplane.com.au/kauer
http://twitter.com/kauer389

GPG fingerprint: A52E F6B9 708B 51C4 85E6 1634 0571 ADF9 3C1C 6A3A
Old fingerprint: E00D 64ED 9C6A 8605 21E0 0ED0 EE64 2BEE CBCB C38B


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Thu, 16 Nov 2017 21:03:07 David Boxall wrote:

> I'd say if the autonomous system performs marginally better than the average 
> driver, then that's a good start. Average includes those who are drunk, 
> drugged, tired, distracted or stupid.

That won't work because you're asking people to place the lives of themselves 
and their families in an opaque piece of technology with the vague assurance 
that, on average, it will be good for the accident statistics.  I don't think 
so!


On Thu, 16 Nov 2017 20:17:27 Karl Auer wrote:

>> If you don't like my benchmark for an autonomous system designed for vehicle 
>> control in any circumstances where a domestic or commercial vehicle would 
>> now operate, would you like to propose one?
> 
> I just think your benchmark is confused - in some ways unreasonably high and 
> in some ways laughably low. But worst, you can't measure it.

The practical question of implementation only becomes relevant when there's 
agreement about the goal.  And the goal needs to be expressed in ethical terms 
because the technology is so opaque and human lives are at stake.  We need to 
put aside the fascinating techie stuff here.


> Easily said. Why don't you outline how a vehicle purporting to meet your 
> benchmark could be tested against it.

In principle, a set of use-cases could be developed which would demonstrate 
whether a vehicle met the benchmark.  For example:

(1)  The vehicle is to be able to climb a (given) mountainous road with no 
safety fence to mark the edge on a rainy night at a speed of at least 'v' kph, 
and must do so successfully on 100 consecutive attempts.  That might test the 
vehicle's sensors.

(2)  The vehicle is to come to a complete stop from a speed of 'x' kph in 'y' 
seconds when an obstruction is suddenly placed 'z' metres in its path, and must 
do so on 100 consecutive attempts.

...that sort of thing.  What else would you expect an expert driver to be able 
to do?  We need another use-case to test the vehicle in an environment where 
there are many vehicles doing erratic things at high speed, for example in a 
multiple vehicle accident where a truck suddenly diverts into an oncoming line 
of traffic.  I'm thinking of the accident on the Hume Highway a little time ago 
when a truck did just that.

David L.

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David Boxall]

On 16/11/2017 5:00 PM, Karl Auer wrote:

... somewhere, somehow, someone was stupid enough ...
I recently had to employ a licenced cabler to run some cat5. I've done 
it hundreds of times, but now it's illegal. There are enough stupid 
people in the world doing enough foolish things that our government 
decided it had to regulate.


Reminds me of the US case where some idiot injured himself with a power 
saw and successfully sued the manufacturer. Now every saw carries a 
warning along the lines "if you run this tool across your hand while the 
blade is spinning, your fingers might fall off".


I have a mower that carries similar warnings - twice. "If you put your 
hand under the deck while the blades are running, you might lose some 
fingers." Then it repeats the warning for feet and toes. I'm only 
surprised that they didn't do the same for the head and genitalia!


 https://www.youtube.com/watch?v=IJNR2EpS0jw

On 16/11/2017 5:59 PM, David wrote:
> ... If you don't like my benchmark... would you like to propose one?
...
I'd say if the autonomous system performs marginally better than the 
average driver, then that's a good start. Average includes those who are 
drunk, drugged, tired, distracted or stupid.


--
David Boxall|  The world is run by
|  those who turn up.
http://david.boxall.id.au   |  --Tony Windsor
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Karl Auer]

On Thu, 2017-11-16 at 17:59 +1100, David wrote:
> If you don't like my benchmark for an autonomous system designed for
> vehicle control in any circumstances where a domestic or commercial
> vehicle would now operate, would you like to propose one?

I just think your benchmark is confused - in some ways unreasonably
high and in some ways laughably low. But worst, you can't measure it.

> That can't be too difficult!

Easily said. Why don't you outline how a vehicle purporting to meet
your benchmark could be tested against it.

Personally I don't think there is any useful benchmark except the death
and injury statistics. Entire classes of error will predictably vanish
- speeding, for example. It is reasonable to assume that new classes of
error will emerge, as unique to autonomous vehicles as drowsy driving
is to humans. Certain forms of analysis paralysis, perhaps. But
probably easier to program our way out of :-)

Regards, K.

-- 
~~~
Karl Auer (ka...@biplane.com.au)
http://www.biplane.com.au/kauer
http://twitter.com/kauer389

GPG fingerprint: A52E F6B9 708B 51C4 85E6 1634 0571 ADF9 3C1C 6A3A
Old fingerprint: E00D 64ED 9C6A 8605 21E0 0ED0 EE64 2BEE CBCB C38B


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Thu, 16 Nov 2017 17:00:47 Karl Auer wrote:

> People are generally far too quick to set up straw men or impossible hurdles 
> for the purpose of attacking such technologies. Most of the arguments I've 
> seen against autonomous vehicles boil down to "I won't be trusting them 
> consarned things until one wins the Paris to Dakar unaided"

"consarned"???

If you don't like my benchmark for an autonomous system designed for vehicle 
control in any circumstances where a domestic or commercial vehicle would now 
operate, would you like to propose one?

That can't be too difficult!

David L,

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Karl Auer]

On Thu, 2017-11-16 at 15:18 +1100, Jim Birch wrote:
> "It was not autonomous, the driver was legally in control."
> That's a legal technicality, isn't it?

No, not really. My old Subaru had cruise control, and in the handbook,
in bold type, was the statement that "engaging cruise control does not
permit the driver to release the steering wheel" or words to that
effect. Such warnings almost always mean that somewhere, somehow,
someone was stupid enough to believe that engaging cruise control meant
the car would steer itself.

If I had let go of the steering wheel after engaging cruise control,
then had a crash, would you say the cruise control had failed?

I'm sure autonomous vehicle control can be improved, will be improved,
and presumably will continue to be improved indefinitely.

People are generally far too quick to set up straw men or impossible
hurdles for the purpose of attacking such technologies. Most of the
arguments I've seen against autonomous vehicles boil down to "I won't
be trusting them consarned things until one wins the Paris to Dakar
unaided".

Regards, K.

-- 
~~~
Karl Auer (ka...@biplane.com.au)
http://www.biplane.com.au/kauer
http://twitter.com/kauer389

GPG fingerprint: A52E F6B9 708B 51C4 85E6 1634 0571 ADF9 3C1C 6A3A
Old fingerprint: E00D 64ED 9C6A 8605 21E0 0ED0 EE64 2BEE CBCB C38B


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David Boxall]

On 16/11/2017 3:18 PM, Jim Birch wrote:

"It was not autonomous, the driver was legally in control."

That's a legal technicality, isn't it?
...
Perhaps, if the manufacturer had ever pretended that the vehicle was 
autonomous. To the contrary, they went to great pains to emphasise that 
the driver should not rely on it. The software wasn't up to that point.


It was (and is) driver-assist technology. Like antilock brakes, 
automatic stability control and collision avoidance systems. Steps in 
that direction, but not autonomy. The manufacturer knew that. The 
manufacturer said it. Repeatedly.


--
David Boxall|  Any given program,
|  when running correctly,
http://david.boxall.id.au   |  is obsolete.
|   --Arthur C. Clarke
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Jim Birch]

"It was not autonomous, the driver was legally in control."

That's a legal technicality, isn't it?

My bottom line would that the autopilot failed in this case (and should,
and presumably was, improved.)

This is one data point.  It proves nothing except that machine car control
is not perfect and the human executive controller failed to monitor the
automatic system sufficiently as the vendor told him to.  This tells us
just about nothing about appropriate public policy wrt driverless vehicles.
Machine control is imperfect; people always don't do what they're told.
Not a lot of useful information there.

To get good public policy we would have to gather some real numbers and do
some comparative analysis.

Jim



On 16 November 2017 at 11:28, David Boxall  wrote:

> On 15/11/2017 10:44 PM, Karl Auer wrote:
>
>> ...
>> Wonder how it would have played out if that truck had been autonomous
>> too. ...
>>
> For a start, the Tesla was in "driver-assist" mode. It was not autonomous,
> the driver was legally in control. The Tesla had radar and computer vision,
> but the software was set up to give priority to the latter. Neither the
> driver nor the computer vision saw the white semi-trailer against a bright
> sky.
>
> The driver relied on the autopilot, even though Tesla warns against it.
> "The car issued six audible warning alerts that he'd spent too long with
> his hands off the wheel."
> https://www.theregister.co.uk/2017/06/20/tesla_death_crash_a
> ccident_report_ntsb/
>
> That is certainly not an _autonomous_ vehicle incident.
>
> --
> David Boxall|  When a distinguished but elderly
> |  scientist states that something is
> http://david.boxall.id.au   |  possible, he is almost certainly
> |  right. When he states that
> |  something is impossible, he is
> |  very probably wrong.
>   --Arthur C. Clarke
>
> ___
> Link mailing list
> Link@mailman.anu.edu.au
> http://mailman.anu.edu.au/mailman/listinfo/link
>
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David Boxall]

On 15/11/2017 10:44 PM, Karl Auer wrote:

...
Wonder how it would have played out if that truck had been autonomous
too. ...
For a start, the Tesla was in "driver-assist" mode. It was not 
autonomous, the driver was legally in control. The Tesla had radar and 
computer vision, but the software was set up to give priority to the 
latter. Neither the driver nor the computer vision saw the white 
semi-trailer against a bright sky.


The driver relied on the autopilot, even though Tesla warns against it. 
"The car issued six audible warning alerts that he'd spent too long with 
his hands off the wheel."

https://www.theregister.co.uk/2017/06/20/tesla_death_crash_accident_report_ntsb/

That is certainly not an _autonomous_ vehicle incident.

--
David Boxall|  When a distinguished but elderly
|  scientist states that something is
http://david.boxall.id.au   |  possible, he is almost certainly
|  right. When he states that
|  something is impossible, he is
|  very probably wrong.
  --Arthur C. Clarke
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Karl Auer]

On Wed, 2017-11-15 at 14:32 +1100, David wrote:
> > This is a complex task, but guess what? Computer systems can
> > already drive cars, and they can do it well.
> Not when they T-bone trucks and kill the driver!

Humans do the same thing - hundreds or thousands of times worldwide
every day and most are ordinary drivers. Not drunk or mad, though no
doubt some are.

It looks to me as if autonomous vehicles are now at about the same
level as human drivers in general. Probably better, because there are
large classes of mistake that they do not make - such as using mobile
phones, driving while fatigued, getting distracted by the kids, driving
drunk... I know you set the bar to exclude those, but that's just
excluding reality.

In the "game" that is partly autonomous traffic, not all questions have
clear answers, and demanding that autonomous vehicles supply them and
never ever fail is pointless and unrealistic. A magic bullet that drops
fatalities and injuries to zero would be nice, but I'll take a
statistically significant improvement.

Wonder how it would have played out if that truck had been autonomous
too. Or if the two vehicles had been able to communicate.

> I'm not proposing any limit, I'm suggesting they have a very, very
> long way to go.  And we might get clues from the nominal computing
> capacity of just a few cubic millimetres of human brain.

Which has very little to do with anything, see above. A $10 computer
these days can reliably beat the vast majority of normal people at
chess. And they build in pauses in those games because otherwise the
speed of defeat would be too demoralising.

Regards, K.

-- 
~~~
Karl Auer (ka...@biplane.com.au)
http://www.biplane.com.au/kauer
http://twitter.com/kauer389

GPG fingerprint: A52E F6B9 708B 51C4 85E6 1634 0571 ADF9 3C1C 6A3A
Old fingerprint: E00D 64ED 9C6A 8605 21E0 0ED0 EE64 2BEE CBCB C38B


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Wed, 15 Nov 2017 15:26:20 Jim Birch wrote:

>>> This is a complex task, but guess what? Computer systems can already drive 
>>> cars, and they can do it well.
>>
>> Not when they T-bone trucks and kill the driver!
>
> Choosing anecdotal evidence seems wrong to me.  Personally I'd prefer a 
> statistical approach.  Expert drivers have killed a lot of people too.  Hey, 
> let's not let them on the roads either.

This report wasn't anecdotal, it was completely factual.

And no valid statistical conclusion can be drawn from one instance of a fatal 
accident involving a Tesla car operating in full autonomous mode when the total 
population of "Tesla cars operating in full autonomous mode" is also so small.

The Guardian article at  
https://www.theguardian.com/science/2016/jun/14/statistically-self-driving-cars-are-about-to-kill-someone-what-happens-next
  begins:  "One hundred million.  That’s the number of miles, on average, that 
it takes a human driver to kill someone in the United States.  It’s also the 
number of miles Tesla’s semi-autonomous ‘Autopilot’ feature had racked up by 
May this year [2016]."  Those facts are easily checked, although it's not clear 
how much human supervision was involved in the "Autopilot" mileage.

But even assuming the Autopilot mileage involved no human supervision and there 
was only one fatality in that period (the Tesla T-boning a truck), assumptions 
favouring the Tesla, it's certainly not clear the Autopilot accident record is 
any better than that of human drivers.

Let's review the situation when there have been 10 or 20 Autopilot fatalities.

David L.


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Jim Birch]

On 15 November 2017 at 14:32, David  wrote:


> > This is a complex task, but guess what? Computer systems can already
> drive cars, and they can do it well.
>
> Not when they T-bone trucks and kill the driver!
>

Choosing anecdotal evidence seems wrong to me.  Personally I'd prefer a
statistical approach.  Expert drivers have killed a lot of people too.
Hey, let's not let them on the roads either.

"Nearly *1.3 million people* die in road crashes each year, on average
3,287 deaths a day. An additional 20-50 million are injured or disabled."

http://asirt.org/initiatives/informing-road-users/road-safety-facts/road-crash-statistics

Driverless vehicles have a massive amount of catching up to do if they are
going to reach parity with humans.  Today's roads are like the open sewers
of early European cities, responsible for an enormous amount of human death
and suffering.  An effort was required to fix the problem but it was done.

Jim






>
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Wed, 15 Nov 2017 13:39:49 Jim Birch wrote:

> Improvement doesn't require perfection.  It only requires replacing things 
> with something better.  If we were all expert, sober, emotionally stable, 
> continuously attentive drivers that would raise the bar for automated 
> vehicles, but clearly we aren't.

The question is "better than what?".

I proposed the benchmark criterion "the same reliability as an expert human 
driver who is not tired, under the affluence of inchohol or other substances, 
and generally in good form".

When autonomous vehicles can meet that benchmark I'll happily go anywhere in 
one.  However we're clearly a very long way from that point when in only May 
last year a "driver" was killed because his Tesla didn't detect a truck turning 
in front of the car in broad daylight.  And this wasn't an experimental model, 
it was bought in a car showroom.

However autonomous vehicles can potentially be very useful in specialised 
situations, such as the bus mentioned earlier which travels only along a 
predefined route at relatively low speed, and I'd expect to see such vehicles 
in normal use very soon.


> An AI system that drives a car doesn't have to be able to do everything a 
> human does, it just has to drive a car.

But it must do so as well as a human with some defined level of driving 
competence, given the places in which it's designed to operate.


> This is a complex task, but guess what? Computer systems can already drive 
> cars, and they can do it well.

Not when they T-bone trucks and kill the driver!


> What for me is the clincher is they get better at it every year.  Why would 
> you presume there is a ceiling to AI improvement, especially one at a level 
> close to current capability?

I'm not proposing any limit, I'm suggesting they have a very, very long way to 
go.  And we might get clues from the nominal computing capacity of just a few 
cubic millimetres of human brain.

David L.

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Jim Birch]

David wrote:


> The prognostications of so-called "elder statesmen" do not have a good
> record.


Especially if you cherry pick them.


>   More recently we had the Prime Minister, an experienced & senior lawyer,
> declare that ...


(Unfortunately for us) politician are in the business of wishful thinking,
and talking.  "Elder statesman" refers to people who have been at the top
but are no longer active so would have accumulated some wisdom but have no
interest in deceiving.  A serving PM would hardly make that category.

A fully autonomous vehicle which can be trusted as a functional replacement
> for a conventional one would have to have the same reliability as an expert
> human driver who is not tired, under the affluence of inchohol or other
> substances, and generally in good form.
>

Improvement doesn't require perfection.  It only requires replacing things
with something better.  If we were all expert, sober, emotionally stable,
continuously attentive drivers that would raise the bar for automated
vehicles, but clearly we aren't.

If this must-be-perfect argument was correct we would not, for example,
allow ABS brakes in cars because they reduce the capabilities of expert
drivers.  However, with the current collection of drivers we have on roads
ABS brakes are a great idea.

Leaving aside the issue of consciousness...
>

Good. Let's stick with functional capabilities.

Furthermore, the brain has evolved its capability over many millions of
> years ... the number of synapses...
>

Evolution produces things that are "good enough" to replicate themselves
into the next generation, statistically not individually.  Perfection was
never part of the brief.  Brains are brilliant at energy conservation.
However, the ways that the brain achieves this feat include a slew of
cognitive biases, perceptual shortcuts, and intermittent attention.

Finally, AI has been around since the early 1960s, when it was confidently
> predicted we'd see human-like intelligence in 10 or 15 years.


Again, there has also been a bunch of people poo-pooing the idea.  As a
matter of historical fact, AI has been improving slower than the boosters
have said, but faster than the naysayers expected.

An AI system that drives a car doesn't have to be able to do everything a
human does, it just has to drive a car.  This is a complex task, but guess
what? Computer systems can already drive cars, and they can do it well.
What for me is the clincher is they get better at it every year.  Why would
you presume there is a ceiling to AI improvement, especially one at a level
close to current capability?

Jim
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[David]

On Mon, 13 Nov 2017 11:49:31 Jim Birch wrote:

> Auto-industry elder statesman declares game over. “It saddens me to say it, 
> but we are approaching the end of the automotive era.  [...]  The end state 
> will be the fully autonomous module [...]

The prognostications of so-called "elder statesmen" do not have a good record.  
Although Thomas Watson probably didn't make the famous statement about a market 
for only five computers, plenty of others made similar predictions at the time. 
 More recently we had the Prime Minister, an experienced & senior lawyer, 
declare that Barnaby Joyce was eligible to sit in Parliament "and the High 
Court will so hold".

A fully autonomous vehicle which can be trusted as a functional replacement for 
a conventional one would have to have the same reliability as an expert human 
driver who is not tired, under the affluence of inchohol or other substances, 
and generally in good form.

Now driving in dificult conditions probably uses a significant part of the 
human brain.  Leaving aside the issue of consciousness and the extent to which 
that might be necessary, the number of synapses in the human brain is estimated 
as about 16*10^8 synapses per cubic millimetre - see
https://academic.oup.com/cercor/article/9/7/722/270779/Estimation-of-the-Number-of-Synapses-in-the

Furthermore, the brain has evolved its capability over many millions of years.

While the neural networks of AI developers may resemble the one between our 
ears in some respects, it's pretty clear they're not within cooee in terms of 
scale & complexity.

I don't think we're going to see any autonomous vehicle with the capability of 
an expert human driver in the forseeable future, despite the Tesla.  But that's 
setting a high bar, and it's easy to see more specialised autonomous vehicles 
like the bus mentioned by Marghanita - see
http://www.latrobe.edu.au/technology-infusion/autonobus

Finally, AI has been around since the early 1960s, when it was confidently 
predicted we'd see human-like intelligence in 10 or 15 years.  I can remember 
my supervisor experimenting with neural networks even then, and sure enough you 
could see it "learning",

David L.


___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Marghanita da Cruz]

On 13/11/17 13:46, Bernard Robertson-Dunn wrote:

On 13/11/2017 11:49 AM, Jim Birch wrote:

Auto-industry elder statesman declares game over. “It saddens me to say it,
but we are approaching the end of the automotive era.

Some, most, all (?) accidents happen because of unexpected circumstances.

What are the chances that automated systems can be built that will deal
with some, most, all (?) unexpected circumstances? And that includes the
software in the automated systems themselves.



As we all know IT systems very rarely end up replacing like for like - 
we need to broaden our thinking on the opportunities and operation of 
automated vehicles.




Ten rules for cities about automated vehicles

Traditional urbanism evolved over millennia to meet human needs. The adoption 
of AVs should not be allowed to replace time-tested places with something that 
would probably make our lives worse.

https://www.cnu.org/publicsquare/2017/10/16/ten-rules-cities-about-automated-vehicles

Marghanita
PS I hated the implication that Pay TV was synonymous with Cable
I now dislike the implication that Electric vehicles are synonymous with 
self driving.

--
Marghanita da Cruz
Telephone: 0414-869202
Email:  marghan...@ramin.com.au
Website: http://ramin.com.au

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Bernard Robertson-Dunn]

On 13/11/2017 11:49 AM, Jim Birch wrote:
> Auto-industry elder statesman declares game over. “It saddens me to say it,
> but we are approaching the end of the automotive era.

Some, most, all (?) accidents happen because of unexpected circumstances.

What are the chances that automated systems can be built that will deal
with some, most, all (?) unexpected circumstances? And that includes the
software in the automated systems themselves.

-- 

Regards
brd

Bernard Robertson-Dunn
Canberra Australia
email: b...@iimetro.com.au
web:   www.drbrd.com
web:   www.problemsfirst.com

___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

Re: [LINK] Cars, again

[Robert Brockway]

On Mon, 13 Nov 2017, Jim Birch wrote:


Auto-industry elder statesman declares game over. ?It saddens me to say it,
but we are approaching the end of the automotive era. Travel will be in


Personally I'd love to have a self-driving car.  Instead of driving I can 
use my time in the car more profitably.


I can even send it on errands.  You've got friends or family visiting from 
interstate but you need to be somewhere else at that time?  Send the car 
to pick them up at the airport!  They can open it with a PIN or voice 
command.  A lot cheaper than a taxi.



standardized modules. The end state will be the fully autonomous module
with no capability for the driver to exercise command. The tipping point
will come when 20 to 30 percent of vehicles are fully autonomous. Countries
will look at the accident statistics and figure out that human drivers are
causing 99.9 percent of the accidents.


Having even 20-30% of cars coordinating their activities will massively 
reduce congestion.  Computer modelling and real-world tests have shown 
this happens with far less than 20% of cars in an area coordinating.


Human driving will live on as a recreational activity permitted within 
designated areas.


Bring it on, I say! :)

Rob
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link

[LINK] Cars, again

[Jim Birch]

Auto-industry elder statesman declares game over. “It saddens me to say it,
but we are approaching the end of the automotive era. Travel will be in
standardized modules. The end state will be the fully autonomous module
with no capability for the driver to exercise command. The tipping point
will come when 20 to 30 percent of vehicles are fully autonomous. Countries
will look at the accident statistics and figure out that human drivers are
causing 99.9 percent of the accidents.

Bob Lutz is a former vice chairman and head of product development at
General Motors. He also held senior executive positions with Ford,
Chrysler, BMW and Opel.

http://www.autonews.com/article/20171105/INDUSTRY_REDESIGNED/171109944/bob-lutz:-kiss-the-good-times-goodbye
___
Link mailing list
Link@mailman.anu.edu.au
http://mailman.anu.edu.au/mailman/listinfo/link