Closed Nikunj-Gupta closed 3 years ago
Suggestion: Removing them. Why? - In Markovian environments, the previous actions should already be a part of the current state. Giving it again explicitly just seems to be redundant and slowing down HAMMER's learning..
Added toggle for it --- can now choose to keep/remove it as and when required.
In CN, we have discrete actions, which we are simply appending to the local observations before passing it to the global agent. We must consider converting them to 0-1 range before concatenating I think. Or, we can also consider removing them, as local observations already contain information regarding positions and velocities of local agents.