[SOLVED] Is there a way to use SciPy's differential evolution, or another library, to minimize one of a multi-output regressor's outs and bound the others?

Is there a way to use SciPy's differential evolution, or another library, to minimize one of a multi-output regressor's outs and bound the others?

I am using a simple RandomForestRegressor from sklearn to predict 3 outputs, and it works well. The problem is that I now need to optimize one of the outputs to be as low as possible, while ensuring the other two predictions don't escape a couple of bounds in the process. I know I could pass the prediction function to something like differential_evolution to minimize the output I need, but my understanding is that you can only enforce bounds on the inputs and not other outputs.

Is there a way to do this that you know of? Am I perhaps approaching this problem incorrectly?

Minimizing the output I wanted results in the other outputs escaping the boundaries I need to set. I'm hoping there is a more intelligent way to find the global minimum other than testing a huge range of inputs, and then finding the minimum among the "allowed" predictions where the other outputs are within the bounds.

Solution

Use something like the following to supply to differential_evolution. Turn the polishing option off.

def my_objective(x):
    y = RandomForestRegressorFun(x)
    if np.logical_or(y[1:] > 2.5, y[1:] < 0).any():
        return np.inf

    return y[0]