Artificial Intelligence

TRACS is currently being used to study computational intelligence at the MITRE Corporation as part of a research project called "Mental Models in Naturalistic Decision Making".

The object is to develop computational models of both cognitive (human) and normative (optimal) strategies in probabilistic reasoning. Computational models of cognitive strategies allow us to test our theories by implementing the models in computer agents (simulations). Then, if the agents perform like people, we will have captured at least some aspects of human thinking in a "usable" model rather than just in a "hand waving" theory.

As one example, the plot on the left shows results from a computational model that was built to simulate how people play the game of Straight TRACS. The dotted line shows how a "perfect agent" would play (normatively) and the solid black line shows how human subjects played (cognitively). The solid gray line shows how a computational model (implemented in a computer agent) played the game. The plot shows good agreement between model (gray) and human (black), both of which vary systematically from perfect performance (dotted). For more details click here.

As another example, a computational model of human thinking was used to program the computer agents (opposing players) that you face in the online version of Poker TRACS. These computer agents are just a first cut, based on a simple model, but most people who play the game find them competitive and humanlike.

Besides being the best test of a cognitive theory, computational models also have many practical applications. An obvious one is for computer gaming - to program computer agents that play like real people so that people can play games even when there are no other people around to play with. A more practical (less recreational) use is to simulate human performance in complex military and industrial "systems" - to save the money and avoid the dangers of doing such testing with real people "in the loop". Finally, computational models of cognitive functions are useful because they provide a direct measure (via mathematical analyses and computer simulations) of how cognitive (human) strategies stack up against normative (optimal) strategies - so engineers can design computerized "systems" to better "support" human thinking. For an example c lick here.

Computer modeling of decision strategies in TRACS is motivated by the landmark research of political scientist Robert Axelrod. In his work, Axelrod simulated the outcomes of various decision strategies in another simple game, called the "Prisoner’s Dilemma”, and thereby gained valuable insight into the development of cooperation in competitive environments:

“… although agent-based modeling employs simulation, it does not aim to provide an accurate representation of a particular empirical application. Instead, the goal is … to enrich our understanding of the fundamental processes that may appear in a variety of applications… For this purpose, realistic representation of many details is unnecessary and even counterproductive … the primary aim is to undertake the exploration in a manner so general that many possible settings could be illuminated.” (Robert Axelrod, “The Complexity of Cooperation: Agent Based Models of Competition and Collaboration", Princeton University Press, 1997).