salvododo
commented
10 months ago Hi, we do have scripts that we use to process the data and convert it into what we could broadly call a trajectory, but there is some IP involved in the scripts and we are not at liberty to share the code.
However, let me give you some pointers. First, note that data is collected individually for each configuration and there is no dynamic update of actions that would give you a "clean trajectory" with well-defined states. So a possible way to construct a trajectory is to connect each current state (which you could represent as the metrics/parameters listed in the CSV file) with a similar state included in a CSV file produced under another action profile. Let's say that you are in state s_t with action a_t, and you want to move to action a_t+1 and need to extract a new state s_t+1. You can construct that state (or, at least, an approximation of the state) by using some similarity scores so that you can pull a synthetic state from the CSV files.
It's not straightforward but this can help you in re-creating the temporal relationship between states.
Do you have any code/scripts or documentation that describes how to use this dataset to do deep reinforcement learning? Is there a way to map this dataset into something that resembles a trajectory of <state, action, reward, next-state> rollouts?