>>>>Whenever someone says they have the software which will handle ALL cases, I generally think it's ALL covered by a try-catch. Well, I don't like the catch.
>>>
>>>In the short term SD cars will require that humans take over in tricky situations. As machines learn those will become fewer and fewer.
>>>
>>>The "mayday" scenario was not for potholes or manhole covers any more than a human would consider those emergencies. As for the rest of those scenarios, ask Google. If you don't like the answer, ask again 6 months from now. Rinse & repeat.
>>
>>That's the trouble. It's
covering scenarios. A finite list of them. When the number of scenarios covered reaches 90% of the statistical sample, the development will stop because there won't be any profit, any new market to be gained by covering the remaining 10%. They'll simply declare victory and move to the more profitable pursuit of cutting costs to the existing models' production, removing features which are too complicated for the 90% of the human observer-drivers and add those features which may prompt them to upgrade to a newer version of the car's OS. Some of which may introduce new ad spaces (last year I rode a cab which was running local ads on its comm/GPS - if you're ever in Zrenjanin, don't ride the cab responding to 588-888).
>>
>>>Many years ago I programmed a CNC mill to carve my initials into a block of acrylic, and saw a pen plotter in action for the first time. I marveled at the inhuman speed and precision.
>>
>>Care to count all the decisions that the plotter had to make in unexpected situations?
>
>You're making a couple of assumptions:
>
>- that a machine's set of responses to scenarios is and always will be a subset of any human's
Not necessarily - it will have to have its own set of problems to solve which are caused by machine driving :). My assumption is that the code for it is written by humans, and that's humans in the west, who generally don't start from a set of axioms and logic, but from a list of cases and scenarios. Want an example? Look at the code for menus. There's a separate set of names for horizontal menus and for vertical menus. The horizontal is called a menu and it has pads; the vertical is called a popup or a dropdown and it has bars. The only difference is orientation, the principle is the same, but the software designers still wrote the code twice. I don't have to imagine how this would work applied to newtonian mechanics - I've already seen it in action.
>- that machine learning can't infer general principles in the same or a superior manner to humans
Machine learning, in that sense, is bound to produce an endless array of false positives, to which it will have to be directed to make exceptions. And unless the cars' OSes merge their collective knowledge (not SkyNet, just CarNet) they'll come up with millions of different rules. No, scratch that, not rules - decision tables.
>About 6 months ago I saw a cool driving tip. If you're waiting to turn left at an intersection, leave your front wheels pointed straight ahead until you make the turn. If you turn them to the left in advance, and then you're struck from behind, you get pushed into the path of oncoming traffic.
Neat :). I've learned a similar one from a cabbie here - get into the crossing (correct this politically, it's an intersection, aka a cut between) as far as getting your front wheels a bit ahead of the median of the road on your left, going straight, so that you make room for more cars behind you.
>SD cars can "relentlessly pursue perfection"; humans, not so much - no motivation, for the vast majority.
I think I'll start that list of cases they'll never get to cover.
