Hi, Ivan,

Thanks for sharing the article. Very good idea of using gradient decent and simulation to tune the PIDs, I am quite inspired.

However, after reading your post, it was not so clear to me what your error function is, I initially thought you are minimizing the “steer” come from the PID output. This seems to be minimum control optimization, which doesn’t make sense. Then I had a look of your code on github, and now I just to confirm with you if I have understood correctly.

The error function to be optimized is f(steer) = cross_track_error² (steer is hidden in this function, since it involves the complex dynamics of the car, but we know the end result, the cross_track_error). Then you optimize the function with respect to Kp, Ki, Kd parameters.

So for example, for optimzing Kp, the gradient from f(steer) with respect to Kp, gonna be (df/dKp) = (df/dsteer)*(dsteer/dP) = dE * dx.

dE and dx are symbols taken from the following gradient decent equation from your code:

void PID::adjust(double &Kx, double dx, double dE)

{

double partialDKx = Kx * dx * dE * learnRate_;

Kx -= partialDKx;

}

Now it is not clear to me why you have Kx, in this case, Kp, also in the partial derivative term?