Partial observability settings

[Santiago Ontanon]

Hi,

We are reconsidering the rules for the "partial observability track". The question is wether the agents should:

1) Be given the initial state of the board as fully observable, or perhaps have a special function where the agents are given some time (e.g. 1 minute) to analyze the map before the game starts. So, they would know the initial state. This would be more similar to the way the StarCraft competitions happen (since the agents do know the map). The difference is that in StarCraft, agents don't know the exact start location of the opponent, whereas in microRTS there is no concept of "start location", and thus the complete start state will be known.

2) Not be given any other information than what the units can see at the beginning of the game.

Any feedback is welcome! :)

[Anderson Tavares]

I think (1) could be more appropriate. It would remind 'Kriegspiel' (or 'partially-observable chess' - https://en.wikipedia.org/wiki/Kriegspiel_(chess)) where we do know the initial state.

I mean, the initial state is known in Kriegspiel, right? :)

[Santiago Ontanon]

Thanks Anderson! I got similar feedback from my student Alberto Uriarte, so, I think I'm going to go that route then!

Let's see if I fid some time later today to update the functionality in microRTS and also the documentation for the competition! 

Thanks again! :)

[Santiago Ontanon]

Ok, this has been implemented!

There is now a function called "preGameAnalysis" in the base AI class, that is optimally called (depending on the tournament settings) before starting a match. This gives the AI the initial state, and also gives the AI time to do any kind of analysis that it wants to do before the game starts.