Adaptive planning in problem-solving: Cost-benefit tradeoffs in bounded rationality.

摘要

This dissertation adopts the bounded rationality framework to study how people adjust the amount of planning in different problem-solving environments. Under the bounded rationality framework, cognition is well adapted to the characteristics of the environment. By exploiting the structure of the environment, complex computations are simplified by efficient mechanisms that may not always lead to optimal solutions, but are sufficient to lead to reasonable levels of performance in various environments with a wide range of characteristics. Simon (1956) used the term "satisficing" to describe this kind of behavior. The bounded rationality framework therefore focuses on outlining the cognitive mechanisms that arise out of the interaction of the environment, the goal of the problem solver, and the bounds of cognition. By studying the tradeoffs between the costs and benefits of planning, the goals of this dissertation are to (i) propose mechanisms for the cost-benefit tradeoffs in adaptive planning, (ii) predict behavior in different problem-solving environments based on these mechanisms, and (iii) provide explanations for sub-optimal performance in certain problem-solving environments.; I began with a Bayesian satisficing model of a general problem-solving situation where the problem-solver has to adjust the amount of planning to improve performance in an uncertain environment. The model predicts that (i) with sufficient experience, the optimal level of performance can be attained, (ii) the responses to changes in costs will be faster than the responses to changes in benefit, and (iii) high planning cost may lead to poor exploration of the problem space. Three experiments were conducted to test these predictions and the results supported the predictions. Cognitive models, built on the ACT-R architecture based on the Bayesian satisficing model, were used to account for the findings in the experiments. The ACT-R models were found to match the empirical data well, suggesting that the mechanisms in the ACT-R architecture were able to explain the adaptive planning behavior observed from the experiments. Implications to learning and performance in various problem-solving environments were discussed.