|| Checking for direct PDF access through Ovid
We study how people terminate their search for information when making decisions in a changing environment. In 3 experiments, differing in the cost of search, participants made a sequence of 2-alternative decisions, based on the information provided by binary cues they could search. Whether limited or extensive search was required to maintain accurate decisions changed across the course of the experiment, but was not indicated to participants. We find large individual differences but that, in general, the extent of search is changed in response to environmental change, and is not necessarily triggered by a reduction in accuracy. We then examine the ability of 4 models to account for individual participant behavior, using a generalization measure that tests model predictions. Two of the models use reinforcement learning, and differ in whether they use error or both error and effort signals to control how many cues are searched. The other 2 models use sequential sampling processes, and differ in the regulatory mechanisms they use to adjust the decision thresholds that control the extent of search. We find that error-based reinforcement learning is usually an inadequate account of behavior, especially when search is costly. We also find evidence in the model predictions for the use of confidence as a regulatory variable. This provides an alternative theoretical approach to balancing error and effort, and highlights the possibility of hierarchical regulatory mechanisms that lead to delayed and abrupt changes in the extent of search.