Prices, e.g., for flight tickets can change almost daily. To minimize the costs, we have to decide when to take an action, i.e.,
when to buy. Such decision tasks are called optimally stopping problems. This paper reconsiders the strongest cognitive
models that are able to predict the average decision maker, adapts them and investigate their predictive accuracy on the individual level, i.e., how good are models in predicting when a participant decides for an action. To perform this analyses, several steps are necessary: (i) Identify data sets that provide raw data for an individual, (ii) develop an individual testing framework to assess the models, (iii) implement and adapt existing models for the individual, and (iv) consider baseline models to assess the goodness-of-fit of the models for the individual. The best and second-best models achieved an overall prediction accuracy of 85% and 84% respectively. Five of the ten examined models managed to beat a strong baseline, proving that they did in fact managed to model the individual decision process.
Reasoning about conditional statements is relevant in science, culture, and our everyday life. It has been shown that humans do deviate from a classical logical interpretation of conditionals. Consequently, in the past years a number of cognitive models based on Bayesian or mental model approaches have been developed, whose performance is normally judged based on their ability to fit aggregate data of participants. Here, we diverge by focusing on the individual instead. Moreover, we propose a different model testing paradigm by analyzing on an existing large data set, how good current models are in predicting an endorsement of an individual reasoner on a scale from 0 to 100%. Towards this goal we reanalyze the data by rigorously distinguishing between test and training data set, by making existing models for conditional reasoning predictable such as the Dual Source Model (Singmann, Klauer, & Beller, 2016) and a model by Oaksford, Chater, and Larkin (2000). We also implement a modeling idea of Pearl based on possible worlds. We can show that all three models perform equally good in predicting an individual reasoner’s endorsement and that they meet an empirical baseline (the median of the most frequent answer). A discussion on the gained insights in understanding conditional reasoning concludes the paper.
Simple laboratory tasks typically allow one or a few methods of task performance. In contrast, moderately complex tasks, such as video games, provide many methods of task performance which, in essence, provide many ways of completing the task without necessarily completing all possible components. Although performance on complex tasks improves with practice, the improvements do not represent the simple effects of power-law learning but, rather, they tend to reflect the discovery and practice of a diverse set of methods. Understanding what we see during complex task learning, requires us to evaluate individual performance against benchmarks of optimality. In this report, we use the game of Space Fortress (SF) as a complex experimental paradigm in which we demonstrate two alternative measures that reveal scopes of individual differences in the discovery and implementation of an optimal method that would be missed by traditional measures of the game.
A complete and holistic understanding of human cognition should be able to relate idiographic parameters representing cognitive functioning to interactions between connected brain networks identified by neuroimaging methods. Here, using the ACT-R cognitive architecture, we examine the possibility of producing idiographic parameterizations of cognitive functioning in a task environment and show that these parameterizations produce reasonable predictions of individual behavior. We then demonstrate a method of determining a subset of parameters that are adequate for prediction of behavior before confirming that the most critical of these task-based parameters is related to functional connectivity measures in individual resting-state fMRI data.