[Python] Replacing differential evolution with Reinforcement Learning for optimized control of...

I am currently optimizing a control problem where I am trying to move a network from random to desired state.

The decision variables are following:

• a vector of n continuous variables
• a vector of m binary variables

At the moment, I am using the standart differential evolution for the continuous variables and a special binary-de algorithm for the binary vector. I would like to know if I could implement a RL model for this setup?

More information: n is usually 10-20, m is usually 80-100. I am converging to some kind of desired state around 10000-20000 diff. evolution tries which takes sometime. Mostly because for each iteration, I need to do some time consuming physical calculations. This calculations can be handled in paralel for the diff. evolution algorithm, however due to licensing, it costs a lot of money. So at the moment, I can do only 1 calculation at a time which takes ca. 1 to 1.5 seconds. In total 10000-30000seconds to finish.

Do you think it is worth while me to learn RL and try to solve this via RL? Please keep in mind RL algo would also need to wait 1.5 seconds for each interaction, since I would send its actions to the licensed software and receive the output of it.

Current Flow:

1. Start algo with initial guesses
2. Send all initial guesses to a blackbox software 1by1
3. Receive all the outputs from the blackbox software 1by1
4. Start differential evolution loop
5. Create new guesses 1by1 send them to blackbox get the output and score them
6. If better score replace the old guess with the new guess
7. do 5-6, 10000 times and get the best score guess

Continue reading...