Description
I want to train agents in partially observable half field offense. However to my knowledge, Half field offense without the full-state flag has the following 2 problems:
1.If I'm correct, HFO uses self_pos which is calculated in base2d code and then translate the relative position of other players and land marks w.r.t this self_pos. This way, HFO uses filtered state_space not observation themselves. For example the agent may have not seen the goal center landmark but because its position is known in advance , our agent has goal center's relative position w.r.t itself because of this self_pos.
2.Agent uses base2d's Turnneck control for its vision. I think we should enable agent to control its own head. So it learns for himself how to control the information flow.
If these issues are plausible, I will be the happiest to contribute. I recommend to create a third feature_set which is solely based on agent's current observation without filtering it. Also add an option to control agent's neck.
Thanks