I'm attempting to convert a standard optimal control problem into a model-based reinforcement learning one. The optimal control problem uses fmincon (or scipy.minimize.optimize if I'm using Python instead of Matlab) to minimize an objective function. I need to replace this optimization step with a zeroth-order optimization algorithm (preferably the cross-entropy method (CEM)). Is there a CEM option in either fmincon or scipy that I'm not seeing? or an equivalent API that I can use instead?
So far I have developed a CEM in Python from scratch to solve this problem, but it is a bit rudimentary and very slow.