A framework for co-ordination and learning among teams of agents

We present a framework for team coordination under incomplete information based on the theory of incomplete information games. When the true distribution of the uncertainty involved is not known in advance, we consider a repeated interaction scenario and show that the agents can learn to estimate this distribution and share their estimations with one another. Over time, as the set of agents' estimations become more accurate, the utility they can achieve approaches the optimal utility when the true distribution is known, while the communication requirement for exchanging the estimations among the agents can be kept to a minimal level.