More algorithm improvements are needed. Right now cubic are rejected (to be used, and jump to their local extreme) because their second derivative at the coordinate of the last found local extreme doesn't match with the second derivative of the fourth degree polynomial fitted around that local extreme; while the rejected cubic doesn't even share a point with the fourth degree polynomial. What I should do is not just store the last_extreme_cubic_ but also the left- or right-most (depending on the direction that we're going into) sample that was used for the fourth degree polynomial fit and the value of the second derivative in that point. And then only reject cubic that use the same point but whose second derivative in that same point deviates too much.

You can follow the progress of the project as a whole here: https://github.com/CarloWood/machine-...
discord:   / discord