Re:Uber shuts down self-driving trucks unit
by andrewksl on Tuesday July 31, 2018 @12:50AM
(#17650617)
> If a truck driver gets in to serious accident every 20 years, an algorithms will be blamed for a serious truck accident every 2.4 months - even if it was 100X better than human! It's an inescapable law of large numbers.
You're comparing 1 driver driving 1 truck to 1 "algorithm" driving 1000 trucks and then comparing the nominal number of accidents.
I'm not an insurer, but I highly doubt they care about individual (human or otherwise) performance. I would bet they do care quite a bit about the accident rates you're attempting to dismiss.
Re:Uber shuts down self-driving trucks unit
by sytelus on Tuesday July 31, 2018 @01:06AM
(#17650664)
Each insurance claim has two sides. Someone is ultimately paying someone else. If one algorithm that controlled 1000 trucks was being blamed for accident because of spurious very low probability mistakes, who do you think will be the payer?
Re:Uber shuts down self-driving trucks unit
by andruby on Tuesday July 31, 2018 @12:45AM
(#17650598)
> Even if autonomous algorithm was 10X better than human, when it takes on responsibility of driving 100 trucks, it would commit 10X more accidents than an average human. It's an inescapable law of probability.
Could you explain this? Are you comparing an algorithm driving 100 trucks to 1 human driving 1 truck?
The single algorithm will not outperform the individual driver if it drives more trucks, but it will outperform the aggregate group of drivers needed to drive the same amount of trucks.
Re:Uber shuts down self-driving trucks unit
by sytelus on Tuesday July 31, 2018 @01:13AM
(#17650695)
My statement is true if each accident by human or algorithm was random error (i.e. not systematic error). When errors are random, you can reduce probability all you want but if you increase number of instances then you are back to square. BTW, this statement is not saying there is increased number of accidents in total. It is saying how many blames gets assigned to an entity. This is the distributed blame assignment problem. When 1000 humans drive 1000 trucks, blames gets distributed. But when same algorithm is deployed in 1000 trucks, each accident gets blamed on same algorithm. So same algorithm assumes all liability that was previously distributed among 1000 humans although total liability in the system might have been reduced.
Re:Uber shuts down self-driving trucks unit
by Erlich_Bachman on Tuesday July 31, 2018 @04:25AM
(#17651475)
The blame is assigned to some entity. But you cannot compare the entity of an algorithm, to an entity of one truck driver - you have to compare to the entity of the collective of all truck drivers.
If you do that, the numbers will favor the algorithm clearly.
The insurance companies are not stupid, they will easily see this logic. It is even in their own economic interest to see this logic.
This is simply a non-argument against automated driving, it does not make sense.
Now consider the fact that each truck carries 100s of thousands of dollars of goods. The wages you pay to driver can easily compensate for the insurance benefits because insurance companies may cover for human mistake but not for the algorithms mistake. Even if autonomous algorithm was 10X better than human, when it takes on responsibility of driving 1000 trucks, it would commit 100X more accidents than an average human. If a truck driver gets in to serious accident every 20 years, an algorithms will be blamed for a serious truck accident every 2.4 months - even if it was 100X better than human! It's an inescapable law of large numbers. (PS: this doesn't mean there are more accidents in total. It means number of blames assigned to same entity is much higher even though overall accident count is smaller).
The autonomous trucking had been attractive only for one reason really: You can avoid last mile. Highways are much more easier for autonomous driving where you rarely have to deal with crossing peds, traffic lights, round abouts etc.
[1] http://atri-online.org/wp-content/uploads/2017/10/ATRI-Opera...
[2] https://www.supplychainconsortium.com/Leadership-Forum/2009/...