Andrew Pohlman wrote: ↑
Thu May 09, 2019 9:38 pm
The AI "brain" of the self driving car is notoriously bad at handling situations it has never seen before. AI must be "trained" via simulations and real road miles. Their pattern recognition is still highly questionable. One of the recent headlines I posted about earlier was that the AI failed to recognize a woman pushing a bicycle across the street and made a decision to mow the "object" over, rather than slow/stop/swerve. The AI can recognize a bicycle, and a woman, but apparently, not a woman pushing a bicycle. This kind of thing is easy for humans. Obviously, not for AI.
Couldn't this be solved with traffic lights?
I see AI as not necessary to solve driverless car.
To me it is more about the brain of the engineers...
Or cost to avoid installing traffic lights everywhere...
I see AI as a current fashion and as such, people tend to overvalue it. I work in Advanced Control for Process Industries. Last year I went to a client to solve a chemical mixing problem. A startup company with specialists in AI used the whole year data and spent 3 months trying to tune an AI to control the process, with not too much success. I solved it in more or less a week only because I knew basic thermodynamics and how to install correctly a proper sensor.
But AI is great when used properly. I am now working with a self-learning MPC solution.