I am training a simple tensorflow model with tensorflowjs that gets 3 variables as its input and generates a quality prediction from them. I'm trying to find a value triple inside the trained model that comes as close to zero as possible (with 0 being a perfect rating). Even better would be if I could find the one zero point that's closest to a given triple so I don't get a different result every time. I've been playing around with a simulated annealing algorithm but I can't get it to work.

Does tensorflow already contain a solution to this problem? Do you have any other ideas how I can make it work?