How come Machine Learning is used in Self Driving cars when we have zero margin for error there?

[Monica Shiralkar]

Machine learning models have a degree of accuracy. This is fine with many areas. However, in case of self driving cars , we cannot have any margin for error as it would lead to accidents. Despite this fact , self driving cars is one of the fields in which machine learning is used. In such a case how is machine learning being used in this field? thanks

[Tim Cooke]

A fine question indeed. I would never wish to put myself in a self driving car.

[Ron McLeod]

In its current state, I wonder how the margin of error for self-driving cars compares to that of humans driving cars?  I would guess self-driving cars are already better and will improve over time (100% chance that I could be wrong).

[Mike Simmons]

Well, I think expecting zero margin of error is, um, unrealistic.  and not something we've required so far before allowing human drivers.  That said, we do need to require extremely small margin of error.   I would say, they need to be demonstrably significantly safer than competent unimpaired human drivers, before they're let loose in public.  But while this is very challenging, I don't see why it should be regarded as impossible, or even unlikely - it's more a matter of time, as far as I'm concerned.  And while it may be a while before public trust develops, I'm confident it will happen.

[Tim Cooke]

From my fairly limited experience with Machine Learning there is one aspect of it that makes me nervous to think about using it for scenarios where an error poses risk to life, and that is the nature of the errors.

Even the better models that I've worked with that operate at a much better average success rate than the corresponding manual process every now and then will go very very wrong, as if it just completely loses any notion of what it's supposed to do, and that in a car scares me.

[Campbell Ritchie]

But real live people driving cars do exactly the same....

[Tim Holloway]

It is a long-known and bitter fact that when automobile manufacturers design a vehicle, they compute how many deaths each feature will cause or prevent and decide cost/benefit for each. As long as we live in an unpredictable world, automobiles, trucks, SUVs, et al. are going to kill people.

If anything, self-driving cars are less likely to kill people. They don't attempt to apply makeup in the rear-view mirror, they don't talk on the phone while driving - or if they do, it's not a distraction at least. They don't experience road-rage. They don't drop  hot greasy fries in their lap and juggle picking them up and grabbing the steering wheel.

However, the operative word there is learning. And remember that virtually all news is based on the unusual. We have vehicular accidents all day every day, but only the unusual ones get special attention. With self-driving cars being relatively new and rare, every accident involving one will get outsized attention, even as the anti-shutdown demonstrators overshadow the much larger crowds of people who don't go out and spread the "love".

As long as vehicles continue to learn, they become progressively safer, allowing for the fact that some situations simply cannot be made safe short of having Superman fly in and lift away a car that cannot find room to stop or swerve. There is after all, a point of diminishing returns and the old "which way to switch a streetcar" philosophical problem has no simple answer.

I think it's generally accepted that traffic would be safer overall if all cars were self-driving. You'd still have external events, but at least then the automated cars could proceed on the basis that all the other cars were being driven rationally.

[Monica Shiralkar]

So, are self driving cars expected to cause lesser accidents than humans driving car?

[fred rosenberger]

So what exactly do you mean by "zero margin for error"?  I mean, does it matter if a self-driving car speeds just a little bit? Does it matter if it cuts somebody off, if that other car (self driving or not) brakes in time?  As others have pointed out, HUMANS don't drive without error, so why would we expect computers to?

[Mike Simmons]

Monica Shiralkar wrote:So, are self driving cars expected to cause lesser accidents than humans driving car?

Yes, exactly.  Now we don't know yet for sure if that will really be possible, but many of us think it will be - and that's the point.  If we can't achieve that, then I think it's very unlikely that self-driving technology will be accepted by the public.

[Bob Winter]

In my eyes not the AI is the problem but us humans reacting unpredictable and all that random stuff we do to the environment the AI has to work in like random constructions to maintain the roads which degrade by normal usage (unless/until someone invents a way to built roads to mitigate that).
As an example: If you have a road not degrading and hence not needing constructions to maintain them you could have it fully packed all by cars all moving at high speeds without any issues when they're all controlled either by a centralized system or using swarm intellegence with eachother. It's proven that a full stop even over a very long distances can be resolved in only seconds if all cars in it start moving at the exact same time with the same speed and acceleration - something a car-to-car AI can do and by this not just help to resolve stuck traffic but to prevent it. It's just one example. Look up "memory of traffic": It just takes one single car to slow down just a bit to produce a full stop several kilometers long and can take hours to days to fully resolve again.
So, it's not the AI itself causing issues - but the fact we try to use only some in a big system without it - that's the human factor.

[Monica Shiralkar]

Mike Simmons wrote:

Monica Shiralkar wrote:So, are self driving cars expected to cause lesser accidents than humans driving car?

Yes, exactly.  Now we don't know yet for sure if that will really be possible, but many of us think it will be - and that's the point.  If we can't achieve that, then I think it's very unlikely that self-driving technology will be accepted by the public.

Thanks