Are you implying that AP learning is close to real time and that quick? What is your theory on how the group leaning works?
I have no idea! All I know is that in two instances (locations) where I decided to do multiple test passes, I had to take control to avoid a serious problem. In the first case, I turned around and went back for another go. I waited until there was no traffic visible in either direction and traversed the same spot at about half my usual speed (to give me time to let it do its worst and not lose control). To my surprise, I went right on by without a problem. So I went back again and tried the spot at regular speed. Again, no problem.
At the next location (same day, same trip), I experienced essentially the same problem. This time I went back and tried it again at regular speed, in case the slow pass changed the outcome and learning. No problem this time either.
For the next few days, I drove past these spots (and others I had taken control but not tested by turning around) with only a very minor steering hesitation that did not require my intervention. Then about a week later I experienced the original problem at the second location. But not since. Ohmman has the knowledge to better speculate on this than I do... I only know what I saw. And whether it was all due to learning in real time or something else, I don't know. I've played with following a car closely and compared to no car ahead and haven't been able to detect a difference. It's beyond my ability to speculate with any degree of certainty, but I find it all very fascinating! (Extremely fascinating - if I had it to do over again, I'd love to get into this area of engineering/software)
There are a number of ways Tesla can be doing this, but first and foremost, it's learning from the driver. That's how these things work, and Elon said as much too. So when you're driving, you're training. When AP is on, it's mostly just obeying the model. There is a slim chance that there is a reinforcement learning algorithm that runs while AP is on, which has a penalty for things like line crossing or a driver taking over without prompting. However, the main training method is for drivers to control the vehicle while it collects environmental data.
Speaking as a non-expert, I would have assumed that a 'reinforcement learning algorithm' would be extremely valuable, especially at the beginning when the system's knowledge is at its lowest. Combining accelerometer data with the driver taking over could take a stab at what was a close call and thus, worthy of further investigation by human minds still working on the software (or prioritized for learning).
One common thought seems to be that the car is learning specific routes or specific locations based on an individual's driving. That's quite unlikely for a few reasons, the most obvious being that a model needs many iterations to adjust. More likely, if a driver is seeing a real difference on the route, the vehicle is learning how to react given a set of inputs which may be unique to that setting. The benefit is that then the model can generalize this to another location on the other side of the country/world where the inputs are similar. And someone there is aiding the model in learning your particular set of inputs. In other words, if AP is doing better on your particular drive, it's because it's doing a better job generalizing to your route.
This makes perfect sense to me. Maintaining a database on every road and curve driven and uploading that data to cars on the routes would be a huge undertaking. However, the parameters must be quite specific, because the problems I was having were with the identical scenario (from the perspective of civil engineer with highway design experience). Same number of lanes. Same tangent (no horizontal curve). Same method of introducing a left turn bay. Essentially the same paint marking design applied to the pavement. Only differences were with the vertical curve parameters. I had an example of each of 'sag', 'crest' and 'none'. The K values were quite large, however, meaning sight distances were good and I wouldn't have expected the learning to be impacted... what was painted on the ground was most important.
Where it appeared to learn the problem at one location, it couldn't apply it to either of the next two. As well, it didn't learn about additional locations on the same route that I had passed several times with me doing the steering... because the problem manifested itself at these locations when I tried AP after a half dozen or more trips with it off. Taking control thus seemed to teach it more than simply driving through multiple times manually.
A simple example is the classic handwritten character recognition problem. If you have a specific way of writing a 4 which isn't classified properly in the existing model, sending it one or two of your 4s with a label ("this is 4") isn't going to do much. However, if there are hundreds of people like you who write slightly similar 4s, and those are part of the labeled training set, it'll properly classify your 4 even without your examples. Compare that to a spot in your route where maybe the line is wavy or worn off at one spot. Surely in the AP collection space, there are other locations like this. Every time a driver completes their trajectory through, the model is getting more confident in that set of inputs. If you waited long enough, the car would work properly even if it hadn't ever seen that particular spot on your route.
It would be interesting to be able to set the fleet learning back to zero and design a few experiments to see all of this in action and improve over time. Too far in to do that now. I'd love to see a detailed technical explanation of the basic logic from Tesla!
