A Unified approach for the design of team decision making structures

In large organizations, decision making is often too complicated to be handled by a single agent or the organizations are divided into smaller functional units. In this situation, a decentralized decision making paradigm is employed frequently. Examples of such organizations include financial institutions, industrial environments, and military command and control systems. There has been a lot of recent interest in the design of optimum decentralized detection structures, e.g., [1-5]. In most of the structures considered, each local detector processes received observations and passes compressed data to the next decision maker as dictated by the network topology. The final decision maker or the fusion center integrates all the received inputs and yields the global or final decision. A variety of configurations such as the parallel fusion network, serial network and tree network have been considered in the literature. The common attribute of all these networks is that the information flows towards the fusion center. Distributed detection systems with feedback have been considered recently [6]. In this structure, information flows from the fusion center to the other decision makers also. In this paper, we provide a unified representation for different decentralized detection network topologies. This representation is inspired by the definition of information structure given in [7]. We first present the definition of the communication structure of organizations for team decision making. A decentralized detection system is assumed to consist of a number of local decision makers and a global decision maker. All decision makers are connected to each other in some fashion. For all interconnection structures, the global decision maker is responsible for making the final decision. The communication structure is specified in terms ofanxn matrix where n is the total number of detectors in a given systems (including the global decision maker). Elements of this matrix take binary values indicating the presence or absence of a communication link between detectors. This matrix can represent many decentralized detection networks such as the serial network, parallel network and tree network. But, it fails to represent the decentralized detection system with feedback. This is due to the fact that time becomes a parameter. Therefore, a generalized communication structure is defined which includes the time parameter t. By properly labeling the matrix with time indices, the generalized communication structure can represent a more general class of decentralized detection structures including the ones with feedback. A number of examples are given to illustrate the use of the communciation matrices to represent various distributed detection structures. The second part of the paper considers the optimum design of any decentralized detection structure represented by its communication matrix under the Bayesian framework. The person-by-person optimization (PBPO) methodology is employed to obtain the decision rules at all the detectors. A number of examples are given to show that the results obtained here reduce to the ones obtained in the literature for specific networks. A fairly complex structure namely a distributed detection system with peer communication is considered to demonstrate the versatility of our unified representation and design methodology.