3 ms·

Sorry, that's just repackaging the problem and pretending it's fixed. The trolley problem is "there's something everywhere inside my stopping distance, we will

Sorry, that's just repackaging the problem and pretending it's fixed. The trolley problem is "there's something everywhere inside my stopping distance, we will inevitably hit something"; yours is only a possible solution for it: "Don't hit objects [evaluated as negative]" with a fallback of "hit the brakes, horn, and do the robot equivalent of pray [ok, gonna hit whatever is in front instead what's front left or front right]". TL;DR: Saying "let's have this one hardcoded solution to the trolley problem" does not make the problem disappear.

corobo 9y ago
It's not supposed to make the problem disappear. The problem doesn't exist frequently enough to do anything with. How much code has to be put in for all of these edge cases? How much data needs to be learned to possibly and nowhere near reliably deal with these edge cases? How often will these edge cases occur? It's not worth the false positives, not by a long shot. You cannot make self driving cars perfect, that's not possible. You just need to make them better than human drivers.
bsder 9y ago
> The trolley problem is "there's something everywhere inside my stopping distance, we will inevitably hit something" You assume humans make correct decisions when faced with that situation. Most of the time humans make incorrect decisions when faced with far less problematic issues. "Swerve" is normally more reflex than thought. The number of times that drivers "swerve" to avoid something like an animal and then crash headfirst into oncoming traffic, thereby injuring humans, is non-trivial.
- D-Coder 9y ago
  "Swerve" ha. I think the typical driver in this situation will scream a bad word, jam on the brakes, and skid into whatever is straight ahead. If an autonomous vehicle does better than this at all, it's a win.
- Piskvorrr 9y ago
  Do I? Nowhere did I bring that up. The TP is a model to compare the possible executions and their outcomes - and yes, human reflexes are not optimized for driving, thus the response is often "crash into other traffic/wayside obstruction", or even "oversteer, flip the car and become briefly airborne". I do agree that "just hit whatever is straight ahead" could be the optimal strategy in many cases, not sure if it makes sense to hardcode it.
TheCoelacanth 9y ago
Swerving seems like it will almost be more dangerous than slamming on the brakes. Never swerve is probably a better heuristic for when to swerve than the one that most human drivers are employing.

corobo 9y ago

It's not supposed to make the problem disappear. The problem doesn't exist frequently enough to do anything with. How much code has to be put in for all of these edge cases? How much data needs to be learned to possibly and nowhere near reliably deal with these edge cases? How often will these edge cases occur? It's not worth the false positives, not by a long shot. You cannot make self driving cars perfect, that's not possible. You just need to make them better than human drivers.

bsder 9y ago

> The trolley problem is "there's something everywhere inside my stopping distance, we will inevitably hit something" You assume humans make correct decisions when faced with that situation. Most of the time humans make incorrect decisions when faced with far less problematic issues. "Swerve" is normally more reflex than thought. The number of times that drivers "swerve" to avoid something like an animal and then crash headfirst into oncoming traffic, thereby injuring humans, is non-trivial.

D-Coder 9y ago
"Swerve" ha. I think the typical driver in this situation will scream a bad word, jam on the brakes, and skid into whatever is straight ahead. If an autonomous vehicle does better than this at all, it's a win.
Piskvorrr 9y ago
Do I? Nowhere did I bring that up. The TP is a model to compare the possible executions and their outcomes - and yes, human reflexes are not optimized for driving, thus the response is often "crash into other traffic/wayside obstruction", or even "oversteer, flip the car and become briefly airborne". I do agree that "just hit whatever is straight ahead" could be the optimal strategy in many cases, not sure if it makes sense to hardcode it.

TheCoelacanth 9y ago

Swerving seems like it will almost be more dangerous than slamming on the brakes. Never swerve is probably a better heuristic for when to swerve than the one that most human drivers are employing.