The main focus of this work is to design a function which, given a prior state of an agent's knowledge about other agents and their observed action(s), returns the new updated state of knowledge.
Our first attempt is to design the belief revision function as a Bayesian update of the state of knowledge represented as a modeling structure defined by the Recursive Modeling Method (RMM). Thus, the function takes the prior modeling structure and the observed behavior(s), and returns a new modeling structure with modeling probabilities updated using the Bayes rule. The fact that RMM's model is a probabilistic estimate of what should be expected of the other agents makes the Bayesian update simple and natural.
See paper Bayesian Belief Update in Multi-Agent Systems for Bayesian update in the Predator-Prey domain.
The simple update function above is, however, just the first step. Additional research issues include the revision of probabilistic information nested deeper in the agent's belief model, and, most importantly, the creation and refinement of the models of the other agents.
MISSION · PUBLICATION · MEMBER · RESEARCH · LINKS
© 1998 Multi-Agent Systems Group at UIC