I am trying to optimize an agent in a simple simulation. For now, I do not want the agent to take the current state into consideration. Instead, it should just react to the environment. My idea is to model the agent as a neuronal network. The simulation results in a score, which I want to maximize.
As far as I understand, I could try to use deep reinforcement learning. However, I have the impression, that it might be faster to treat the simulation as a black box function which I want to optimize. Therefore, I need an optimization algorithm which does not require a gradient to optimize the weights of my neuronal network. If I am not mistaken, deep neuro evolution might be an option. Unfortunately, I haven’t been able to find a simple to use library yet.
Does anybody know of a library which could help me to solve this problem? Right now, it’s most important that it is simple to use and does not require a lot of knowledge.
I am using Python, Keras ans Tensorflow.