Bayesian repeated normal form games.

[gmweinberg]

I was thinking of very abstract game where you can feed in any normal form game and number of repetitions. It looks
like most of the functionality for this already exists, there's already a "normal form" game class and a game transform to make it a repeated game. The twist is, I want each player to only start off knowing its own payoffs and have to (possibly) infer the payoffs of the other players from their actions as the game progresses. I figured the way to do that is to make each players' own payoffs to be part of the observation tensor. Incidentally, there's a little more information I would want to add: right now it has a fixed number of turns of recall, I would like to add a cumulative sum of the number of times a player has played an action with a decay parameter so recent actions appear more important.

The most straightforward thing to do would just be to add some extra optional parameters to repeated game, but I was thinking maybe it's better to leave that alone and create a custom observer. I couldn't find any examples of a custom observer, so maybe adding an example of one would be useful. I don;t know where to put it if I were to create one, or if maybe it's just not a good idea at all.