Decision Strategies in Simulated Environments

Abstract

Simulated task environments resembling medical decision problems and strategies for their solution are investigated. The tasks, which are represented within a computer, contain hypothetical symptoms, diseases, laboratory tests, and treatments, as well as their probabilistic interrelationships. Our objectives are to develop and test strategies for diagnosis and treatment of the diseases. These strategies are also implemented on a computer, and their performance in the medical decision task is evaluated. Within specific tasks, three strategies are examined: (a) an expected utility maximizer, using Bayes' theorem to combine new data with previous observations; (b) a heuristic strategy that searches for satisfactory solutions using informal rules; and (c) a generate-and-test strategy that attempts solutions by using random, trial-and-error searches. The results illustrate the trade-off between decision quality and rule complexity. Potential advantages of simple decision strategies are discussed in a cost-benefit context. Furthermore, the role of knowledge in decision making is also discussed, and the need for explicit models of inductive learning is emphasized. Finally, general implications and possible extensions are noted.