We view the agent as already maintain probabilistic models of other agents in its knowledge base. If a model is not accurate, its probability will gradually decrease when the agent performs probabilistic update based on its observation on other agents. This signals the need for a better model, which can be obtained by learning.

In most situations, information about other agents only come from our observation of their behavior. Let us define a history of an agent's behavior as a set of its observed actions during a particular time frame, in which the data of the world states are known. Given only a history of behavior, how do we learn a better model?

Our idea is to use influence diagrams to model agents. As a modeling representation tool, influence diagram is able to express an agent's capabilities, preferences, and beliefs, which are required if we want to predict the agent's behavior. Learning is done by refining parameters in influence diagram that are associated with capabilities, preferences and beliefs of other agent, based on the history of its behavior. There can be a number of modified models that can be generated by the learning method. We will allow the modified models to compete with each other and other models maintained by the learning agent. The probability of each model being correct can be arrived based on how well the model predicts the history of behavior.

We have developed a method for learning the capabilities and preferences. Here is the paper which describe them:

Dicky Suryadi and Piotr J. Gmytrasiewicz. Learning Models of Other Agents using Influence Diagram, Submitted for publication, 1998.

In our future work we will develop the learning algorithm for modifying another agent's beliefs, and we will integrate all of the learning algorithms with the agent's probabilistic frame-based knowledge base.

*** Last updated on Nov 24, 2001