Is there an RL approach/algorithm that would be suited for the following kind of problem?
- There is a continuous action space with an action value $A_{a,t}$ for each action dimension $a$.
- The objective function is a non-linear function of the satisfaction factors $S_{s}$ for each satisfaction dimension $s$ and some other random & independent factors. This objective function can be known to the agent.
- Each satisfaction factor depends on an independent variable $\Delta^S_{s,t}$ and the effect $\delta^S_{a,s,t}$ of each action: $S_{s,t}=\Delta^S_{s,t} +\sum_a A_{a,t} * \delta^S_{a,s,t}$.
- Each action can further have an effect $\delta^R_{a,r,t}$ on the inventory factors $I_{r,t}$ for each resource dimension $r$, with inventories being kept between time-steps and a factor $\Delta^R_{r,t}$ that is added or removed from the inventory at each step independent of the actions: $I_{r,t+1}=I_{r,t}+\Delta^R_{r,t} + \sum_a A_{a,t} * \delta^R_{a,r,t}$
- The agent is constrained by each of these resources (i.e. the inventory has to remain positive).
- The agent should be able to deal both with $\delta$ and $\Delta$ factors that are visible (states) and invisible (have to be learned).
- A trained agent should be able to know how to adapt to changes of the $\delta$ and $\Delta$ factors, as well as the introduction or removal of activity dimensions.
EDIT: I have adapted the problem description after some feedback.