Could we consider input and output on the decision variables on a feedback problem such as RL?
Could we consider "state" as "input" instead of "input from environment:D"?
Could we consider "state" as "output" instead of "real onput:O"?
My background is on ML. I need to translate a RL techniques working based on ML to mathematical optimization.
For instance:
In a decision making of a optimization problem, during time steps of running, a variable(D) is observed and based on state which is a pair of (D,A,B,C,O), decision about updating the variable (O) is made. Then, this variable (O) helps to update the parameters (A,B,C)by some formula Then cost is updated. the cost function on each time step is defined based on (A,B,O).This problem continues. The goal is to minimize cumulative cost.
What are the input/output/decision variables of the problem if we want to write as an mathematical optimization one?
input: "D" or "state=(D,A,B,C,O)".
output: "O" or "(O,A,B,C)"
decision variable: "state=(D,A,B,C,O)" or"(A,B,C,O)" or "(A,B,C)"
Objective function is to minimize cumulative cost or "cost"?