Communication management using abstraction in distributed bayesian networks

Techniques were developed in previous work for managing communication in a controlled satisficing manner in two layer distributed Bayesian Networks. DEC-MDPs were used to sequence the information transferred in order to guarantee the required confidence level. In this paper, we introduce multiple abstraction layers into the Distributed Bayesian Network as a way of carrying more useful information in transmitted data to further reduce the number of messages that need to be sent. An algorithm is developed to automatically generate appropriate abstraction data. Techniques are introduced to effectively incorporate this abstraction data set into the DEC-MDP framework. We show that the appropriate addition of abstraction data actions simplifies the DEC-MDP while reducing the expected communication cost. This work provides us with a formal view of the use of abstraction in agent cooperation and begins to give us an understanding of when the less abstract data needs to be transmitted.