Imagine we have 2 air conditioner systems (AA) and 2 "free cooling" systems which mix external and internal air (FC) in a closed box which always tends to warm up. For each system, we have to find turn on and off temperatures (for some hysteresis, let's say between the range 20-40 each one) to optimize the energy consumption.
As we don't know the relation between these parameters and the energy consumption (and we don't intend to know them), we treat the problem as a black-box function.
Till now, the problem would be solvable via a bayesian optimizer (eg. with gaussian process acquisition function).
But there is a problem: the best configuration may change between seasons, and even days! A simple bayesian optimizer maybe could deal with these changes limiting the data it takes into account by, for example, the last 15-30 days. But this would deal with the change AFTER the consumption increased.
So, the idea is introduce some contextual variables which would help the system prevent these changes (eg. the external and internal temperature, and the vectors of variation of external and/or internal temperature, the weather prediction, whatever).
Also, some of these variables we can take into account might be internal of the system, which means while these influence the best configuration, the actual configuration also influences these variables! and this becomes a reinforcement learning problem.
1) Is there a way (documented or experimental) to know which variables (both internal or external) influences the optimal configuration of these AA/FC systems?
2) Based on the first question, which would be the best approach?
2.1.) No features. This might be considered a multiarmed bandit problem for continuous reward. (FIX POSTERIOR TO SCENARIO CHANGE, IF THERE IS A SCENARIO CHAGNE)
2.2.) Only external features to predict the scenario change. This might be considered a contextual multiarmed bandit-problem. (FORESEE THE SCENARIO CHANGE)
2.3.) Consider only system-internal features. This can be considered a reinforcement-learning problem. (FIX IMMEDIATLY THE SCENARIO CHANGE)
2.4.) Consider both external and internal features. This can be considered a reinforcement-learning problem where some of the states are not influenced by the configuration. (FORESEE THE SCENARIO CHANGE, AND IF SOMETHING FAILS, FIX IMMEDIATLY).