[seiji]

I think you're over-nerding it here.

Google's cars use Bayesian reasoning. — How does that change anything? What's an alternative to "bayesian reasoning?"

Google can probably tap into the government data for new stop sign installments anyway — such a thing (an accurate government database of new stop signs) does not exist.

the prior probability of a stop signing being there — scenario: a construction worker with a rotating STOP/SLOW hand sign during road work. Or, manual traffic overrides with police standing in the road with hands up or hands waving to manually control traffic.

the car can look at cars in front — assumes said cars exist and the motion of forwardness is more important than any potential cross traffic being stopped for.

[NhanH]

I think he meant that instead of following deterministic rule (ie stop sign == stop), the car can take into account other factors, and make the actions with the best probability to not cause a disaster.

And keep in mind that the world in which self-driving cars are dominant is very and much different than the world we're living right now (as far as transportation are concern). Stop signs are meant to be read for human, car don't actually have to care about it. How about instead of stop sign at crossroad, we have some devices that can monitor the incoming traffic, and relay that information to the car? (I'd say that the car can even communicate with each other by themselves - but that approach has a host of other problems too).

Honestly, I'd say the main problem in a world of self-driving car are other human-driven cars: the self-driving car has no idea what the human would do, and the human has no idea what the automated car would do.