The Project Thesis

"Mental Models" is a MITRE-Sponsored Research project that began in the year 2000 (Roman MM). The project is a blend of Natural Computation and Artificial Intelligence, to understand how people think and to develop computer systems that can improve Command and Control. The yin and yang of it are symbolized by the circle and square.

Natural Computation (Richards, 1988) is the study of how people exploit lawful regularities of the world in order to make useful inferences about it. For example, in vision people readily infer 3-D properties from 2-D images, which is rather remarkable since a given 2-D image can arise from an infinity of 3-D structures. Of course sometimes people's inferences are wrong and the study of these perceptual "illusions" can help us understand how human vision works (Burns, 2001). Likewise, the study of decision making "errors" can help us understand how people achieve situation awareness and select courses of action in Command and Control applications (Burns, 2000).

Artificial Intelligence is the study of how to endow machines with human abilities. By building computational models of human intelligence, scientists can test their theories and engineers can build new systems to help people make better decisions.

Our target is to improve human-system performance in complex domains like military command, intelligence analysis and air traffic control. Our products are computational models of human thinking that can provide a scientific foundation for engineering applications. The main application to MITRE's Technology Program is in the area of "Decision Support", since cognitive models are needed to guide the design of computer systems that augment cognition. Another application is in the area of "Modeling and Simulation", since models of people are needed along with models of systems (like radars, networks, etc.) for training humans with simulated teammates and for testing systems with simulated users. A final application is "Intelligent Information Processing", where computational models of cognitive strategies can be implemented in computer systems as autonomous agents.

The question is not just "how" do people think, but "how well" do people think. To model how and to measure how well, we adopt the normative framework of "Bayesian inference" (Knill & Richards, 1996).

The approach is blend of Behavioral Science and Decision Theory. As an example, Edwards & Phillips (1964) showed how Bayesian mathematics can provide the normative standard for identifying when people are smart and where people can use help in Command and Control situations, so engineers can design support systems like a "Probabilistic Information Processing system" (PIPs).

Perfect rationality requires infinite resources, which is not reality. So the basic theory is one of "bounded rationality" (Simon, 1957), which holds that people are "satisficing" (good enough) rather than optimal decision makers. With this theory, our challenge is to understand the natural constraints that the world imposes on people as "laws", and that people impose on themselves as "wants", since these constraints determine what is "good enough" in decision makers' minds.

The approach starts by assuming that people are Bayesian, at least to a first approximation (Knill & Richards, 1996), and then examines the ways that people deviate from Bayesian standards (Edwards, 1961; 1954). In this "Bounded Bayesian" approach, we pay special attention to the "heuristic" strategies that people employ as short cuts because these heuristics are both the strength and weakness of human thinking (Gigerenzer & Todd, 1999; Kahneman, Slovic & Tversky, 1982).

In an initial case study (Burns, 2000), the Bounded Bayesian approach was used to analyze an infamous "error" in Command and Control. The analysis showed that this error was actually rational (Bayesian) given the information that was available to the decision maker at the time along with his finite resources for gathering and processing additional information. This case study is an example of how a Bounded Bayesian approach can shed light on human thinking and help to improve system designs in Command and Control applications.

Burns, K. (2000). Mental Models and Normal Errors. Proceedings of 5th Conference on Naturalistic Decision Making. Also in Montgomery, H., Lipshitz, R. and Brehmer, B., Eds. (2004). How Professionals Make Decisions (Lawrence Erlbaum).

Burns, K. (2001). Mental Models of Line Drawings. Perception, 30, 1249-1261.

Edwards, W. (1954). The Theory of Decision Making. Psychological Bulletin, 41, 380-417.

Edwards, W. (1961). Behavioral Decision Theory. Annual Review of Psychology, 12, 473-498.

Edwards, W., & Phillips, L. (1964). Man as Transducer for Probabilities in Bayesian Command and Control Systems. In Shelly, M. & Bryan, G. (Eds.). Human Judgments and Optimality (Wiley).

Gigerenzer, G., & Todd, P. (1999). Simple Heuristics that Make Us Smart (Oxford University Press).

Kahneman, D., Slovic, P. & Tversky, A. (1982). Judgment Under Uncertainty: Heuristics and Biases (Cambridge University Press).

Knill, D., & Richards, W. (1996). Perception as Bayesian Inference (Cambridge University Press).

Richards, W. (1988). Natural Computation (MIT Press).

Simon, H. A. (1957). Models of Man (Wiley).