The method we propose is an alternative to traditional approaches to multi-agent systems that rely on the designer's insight into the kinds of interactions the agents will be involved in, and encode it in prespecified protocols. Our is based on decision theory and gives the agents themselves the initiative and the ability to make decisions, based on their up-to-date local information about the current state of the multi-agent environment. As such, RMM is especially viable in domains characterized by unanticipated change, even when the team changes and the communication channels fail in ways that may be contrary to what the designer would have expected.
The centerpiece of our framework is the representation and reasoning with all of the relevant information that an agent may have about the environment, itself, and other agents, necessary to make the rational decision in a multi-agent situation. This has to include modeling other agents to predict their actions. The fact that other agents could take a similar approach gives rise to the recursive nesting of models. RMM proposes a dynamic programming solution of the resulting hierarchy of nested models. Further, it allows the agents to compute the values of various kinds of messages they may transmit to other agents. By choosing the message with the greatest value, the agents can communicate the most essential information when the communication bandwidth is severely restricted, the communication channel could be unreliable, and when the messages could be intercepted.