skopt.gbrt_minimize (gradient boosted regression trees) – how does it calculate the gradient of the loss function if it is not differentiable?

I am optimising 11 parameters for a Jansen & Rit whole-brain model as follows:

skopt.gbrt_minimize – how does it calculate the gradient of the loss function if it is not differentiable?

I am optimising 11 parameters for a Jansen & Rit whole-brain model as follows: