Consider the following problem, simplified:
Let's say that you have some data where you have only 5 columns/attributes.
One of them can be seen as a reward. How well we performed at a given run.
The other two attributes, let's say x and y are our input to the system. A human hand picked those x, y values and they were recorded.
Another two attributes are observations from sensors. These are not inputs and we have no control over them. Let's call them ob1 and ob2.
All are real values (nothing is discrete).
One idea is to look at this dataset as an ordered dataset.
We are asked to choose input values x and y that maximize our reward.
We have available ~70.000 instances of this five-tupled dataset
This is one approach that comes into mind but not sure if it is correct in principle or the simpler one.
We could build a predictor which takes as input the four attributes (x, y, ob1, ob2) and has as target the reward.
Then try a repetitive process for the inputs by using reinfocement learning (?) to get the inputs that maximize the reward?..