This paper is a progress report of the research done in this direction. Its aim is to give a broad view of the ideas that arose during this research, even if they are speculative in nature and not supported with enough evidence. There are three hypotheses emerging from this work: (1) Newell and Simon's Theory of Problem Solving may provide a sufficient theoretical framework for the research on human judgment under uncertainty. (2) Process tracing methods, and especially protocol analysis have the power of providing data for such framework. (3) Bayesian belief networks are a promising candidate for a symbolic representation of problems involving uncertainty.
A preliminary analysis of verbal protocols of over 20 subjects working on various judgmental tasks was performed. A computational framework for modeling judgmental processes was created in which subjects' relevant knowledge was modeled by a large, unique for each subject belief network. When structuring the problems, a majority of subjects tended to instantiate nodes of that network and create deterministic scenarios of possible events influencing the event in focus. These scenarios resulted in problem spaces resembling incomplete probability trees. Computation of uncertainty seemed to be weighting of subjective likelihood of the considered scenarios. Of the two formal mechanisms for probabilistic reasoning within belief networks: belief propagation and generation of deterministic scenarios, the approach observed clearly resembled the latter one. The paper suggests several possible mechanisms that can account for discrepancies between human judgment and the probability theory.