Measuring information transmission for team decision making

A fundamental problem in designing multiagent systems is to select algorithms that make correct group decisions effectively. Typically, each individual in a group has private, relevant information and making a correct group decision requires that private information be communicated. When there is limited communication bandwidth or potential for delays in communication, it is important to select the algorithm for making group decisions that requires least communication. This thesis makes three contributions to the design of multiagent systems. First, it shows the benefits of quantifying information transmitted by measuring the entropy of messages to find algorithms for decision making that minimize use of bandwidth. Second, it provides an analysis of the information content of a diverse group of center-based algorithms, including several types of auctions, for making group decisions. Third, it defines a new data structure, the dialogue tree, that compactly represents complex interactions between individuals. The thesis demonstrates that the amount of communication required by an algorithm is highly dependent on factors of the multiagent system's environment, such as team size, error tolerance, and the likelihood that a given agent can perform a particular task. No single algorithm guarantees the least communication in all environments. The thesis further shows that a system designer must consider both coordination and revelation when choosing an algorithm, and it provides compelling evidence that systems that implement an unsuitable algorithm for decision making incur significant costs for wasted communication.