Distributed decision-making by a team of neural networks

An increasing number of application areas in engineering and computing systems is requiring collaborative interaction between physically distributed decisionmakers and controllers. Such distributed applications frequently give rise to nonlinear distributed decision and control problems in the presence of uncertainties such as the effects of the local decisions on the overall objective and disparate system state information, Artificial neural networks in the past have proved effective in adaptive realization of nonlinear decisionmaking and control rules. In order to apply them to distributed applications, new interconnection models as well as adaptation and learning methods are needed to cope with distributed sources of uncertainty such as those mentioned above. In this paper, motivated by studies in large-scale systems theory, we present a team theoretic interconnection of neural networks. The problem that arises due to the disparate nature of the state information available to the decision-makers is highlighted. This problem is overcome by constructing local approximations to the overall performance function using only the locally available information. These local performance functions (termed local critics), in turn, yield adaptive methods for realization of distributed decision rules using neural networks. Simulation studies have demonstrated the applicability of the above approach to many distributed decision-making problems.