Synergy in teams with incomplete information.

Abstract

A team is collection of agents, i.e., units that autonomously make decisions, committed to act in pursuit of a single objective. Synergy is the ability of a team of agents to exceed the effects of its constituents. If synergy is disrupted, it is said to be brittle. A formal definition of synergy does not exist, preventing a unified study of synergy. Moreover, attaining synergy in teams with incomplete information is a great challenge, and only few solution methods are known. Finally, synergy brittleness is undocumented in team formalisms. This dissertation derives optimal strategies for attaining synergy, identifies when synergy is brittle, and develops methods for exploiting synergy brittleness. First, a formal analytic measure of synergy is defined. The definition allows quantification of the concept and is important to the unification of related research. Moreover, because of its analytic nature, the quantitative measure of synergy can be used for both analysis and design. Second, maximum synergy is characterized in a number of team decision problems. A compact, generalized causality controller is derived for an optimal team decision problem and used to design networked control systems. Linear policies are shown to be optimal for a minimax team decision problem, which expands the class of tractable team decision problems, and makes synergy computable using established numerical algorithms. Third, the classical team theoretic framework is extended to consider inconsistent information. Inconsistent information creates a paradox, i.e., all attempts to model the inconsistencies will, as a result, eliminate them. In the extended framework, the paradox is circumvented through the derivation of dominant beliefs that yield a systematic, optimal way of making decisions despite imperfect models of the world. Fourth, a theory of synergy brittleness comprising a set of necessary and sufficient conditions is developed. Synergy brittleness results when agents disuse information without loss of optimality. This inefficiency is exposed by the application of the necessary and sufficient conditions. Finally, the theory of synergy brittleness is exploited for scalability and for distribution of the computations. Synergy brittleness is shown to decompose the optimization problem. The decomposition is computable using a distributed and efficient cliquing algorithm.