Ramblings on unsupervised learning theory in autonomous cars

I am keen on just talking freely on the speculations and ramblings regarding philosophy of AI, the way it learns.

In Computer Science, in general — the classical algorithms, there is a concept called ‘on-line learning’. It basically means a algorithm has to take decisions, as it goes on, one step at a time; the algorithm does not have a “global visibility” of data. It receives data, piece by piece, just like you could image walking in a unknown city without a map — and you would be basically operating alone, and were not able to ask questions for way-finding.

Now think about the way in which a car’s machine learning would learn to drive. Supervised learning meant that the model gives predictions (choices), and the choice’s “distance” to correct answer can be rather easily calculated. Now this subtraction — the error term, is used to adjust the algorithm’s parameters, called “weights” in a neural net.

We could also do unsupervised learning, in fact. I mean the idea is quite simple, the car can reliably be made to detect, did the car run into a trouble:

• a major trouble like a crash
• off-road, was forced to drive to a uncomfortable place
• car hit some object it was not supposed to hit

The car could thus be made to supervise its own “recently past” learning.

Now, given that a car supervises itself: an interesting question will be raised right here:

If a autonomous AI-car crashes, how does it know whether to blame the crash on its own model (algorithm), or circumstances?

a) how does unsupervised learning know, whether the crash could have been avoided, at all? Even though the car had done the best of best moves, was it completely in-avoidable crash, or not? Ie. was it the car’s “fault”, or just bad circumstances?

b) if the learning would be to be used to teach future cars by sharing the knowledge, exactly when (at what point in time) did the algorithm start doing a sequence of wrong moves?