Bayesian Agencies in Control

Abstract

Most agent architectures implement autonomous agents that use extensive interaction protocols and social laws to control interactions in order to ensure that the correct behaviors result during run-time. These agents, organized into multi-agent systems in which all agents adhere to predefined interaction protocols, are well suited to the analysis, design and implementation of complex systems in environments where it is possible to predict interactions during the analysis and design phases. In these multi-agent systems, intelligence resides in individual autonomous agents, rather than in the collective behavior of the individual agents. These agents are commonly referred to as “next-generation” or intelligent components, which are difficult to implement using current component-based architectures. In most distributed environments, such as the Internet, it is not possible to predict interactions during analysis and design. For a complex system to be able to adapt in such an uncertain and non-deterministic environment, we propose the use of agencies, consisting of simple agents, which use probabilistic reasoning to adapt to their environment. Our agents collectively implement distributed Bayesian networks, used by the agencies to control behaviors in response to environmental states. Each agency is responsible for one or more behaviors, and the agencies are structured into heterarchies according to the topology of the underlying Bayesian networks. We refer to our agents and agencies as “Bayesian agents” and “Bayesian agencies.”