Game theory provides a basis for agent behavior in interactive
situations. 
Given a description of a game, it specifies conditions for rational
behavior.
However, there is no room in classical game theory for agents to
reason about the possibly irrational strategies of other agents, or
about their possibly incorrect assumptions about the world.
It is well known from research in behavioral economics that humans do
not generally play as game theory dictates.
So if we want to build agents that can interact successfully with
humans, or with other imperfect agents, we need to explicitly model
agents' reasoning strategies and thought processes.

This research seeks to address this need by providing a new
representation language, called networks of influence diagrams.
As the name suggests, a model in our language consists of 
smaller models, called multi-agent influence diagrams, linked together
into a network.  Each multi-agent influence diagram 
is a possible mental model of an agent, representing a possible
version of the game and the strategies being used.
Our language allows us directly to model "I think that you think that
I think ..." reasoning, and also allows uncertainty ("I think that maybe
you think ...").
We have looked into learning the parameters of networks of influence
diagrams from data, and are currently looking into modeling how humans
play a negotiation game.

<u>
<li>
<a href = "../Papers/nids.aamas03.ps">
A Language for Modeling Agents' Decision Making Processes in Games,</a>
Y. Gal and A. Pfeffer,
<i>Second International Conference on Autonomous Agents and Multi-Agent
Systems</i>, 
Melbourne, Australia, July 2003.
</li>

<u>
<li>
<a href = "../Papers/nids.aamas03w.ps">
A Language for Opponent Modeling in Repeated Games,</a>
Y. Gal and A. Pfeffer,
<i>Fifth Workshop on Game Theoretic and Decision Theoretic Agents at
AAMAS,
Melbourne, Australia, July 2003.
</li>