LEARNING IN DYNAMIC DECISION TASKS - COMPUTATIONAL MODEL AND EMPIRICAL-EVIDENCE

Dynamic decision tasks include important activities such as stock trad ing, air traffic control, and managing continuous production processes . In these tasks, decision makers may use outcome feedback to learn to improve their performance ''on-line'' as they participate in the task . We have developed a computational formulation to model this learning . Our formulation assumes that decision makers acquire two types of kn owledge: (1) How their actions affect outcomes; (2) Which actions to t ake to achieve desired outcomes. Our formulation further assumes that fundamental aspects of the acquisition of these two types of knowledge can be captured by two parallel distributed processing (neural networ k) models placed in series. To test our formulation, we instantiate it to learn the Sugar Production Factory (Stanley, Mathews, Buss, & Kotl er-Cope, Quart. J. Exp. Psychol., 1989) and then apply its predictions to a human subjects experiment. Our formulation provides a good accou nt of human decision makers' performance during training and two tests of subsequent ability to generalize: (1) answering questions about wh ich actions to take to achieve goals that were not encountered in trai ning; and (2) a new round of performance in the task using one of thes e new goals. Our formulation provides a less complete account of decis ion makers' ability after training to predict how prespecified actions affect the factory's performance. Overall, our formulation represents an important step toward a process theory of how decision makers lear n on-line from outcome feedback in dynamic decision tasks. (C) 1997 Ac ademic Press.