The idea was for a policy to save its own frequency of actions during a simulation (e.g. "Flood Control (5%)"). However this seems to create a number of problems:
-
During parallelization, the policies themselves need to be returned to master because they are modified when running the objective function
-
After snapshots are saved, if a policy is run again in simulation mode, it does not clear the action count. This results in frequencies that sum to more than 100%. Should be an easy fix.
-
The original goal was to prune actions that never occur. But I never got around to doing that anyway.
I'm wondering if it would be easier to just remove this altogether. Action counts are not a property of the policy anyway, they depend on the forcing scenario. Might be better to leave this up to the simulation model to compute, outside of ptreeopt.
The idea was for a policy to save its own frequency of actions during a simulation (e.g. "Flood Control (5%)"). However this seems to create a number of problems:
During parallelization, the policies themselves need to be returned to master because they are modified when running the objective function
After snapshots are saved, if a policy is run again in simulation mode, it does not clear the action count. This results in frequencies that sum to more than 100%. Should be an easy fix.
The original goal was to prune actions that never occur. But I never got around to doing that anyway.
I'm wondering if it would be easier to just remove this altogether. Action counts are not a property of the policy anyway, they depend on the forcing scenario. Might be better to leave this up to the simulation model to compute, outside of ptreeopt.