Learning and Strategy Selection in Probabilistic Environments

Many decisions have to be made on the basis of knowledge about correlational structures in the environment. It has been found that people with a low working memory capacity perform better in a covariation detection task (Kareev, Lieberman, & Lev, 1997). This has been attributed to the assumption that they can only consider smaller samples which are more likely to bear a correlation parameter that exceeds the population parameter. Our data and results from a reinforcement learning model on an extended version of their task do not clearly support this account. As alternative explanations differences in reinforcement learning, hypothesis generation and strategy selection are considered. A very simple strategy (payoff maximization) is most successful in this environment. Differences in capacity and strategy selection are to be further studied in a cue based social categorization task, namely to predict political party preferences. Here, people have to apply knowledge about correlations acquired in the real world. It is hypothesized that people with a lower capacity use simpler strategies that could be, again, even more successful than more complex strategies because they exploit the structure of the environment. This could be reflected in behavioral data and in differential model fit to a variety of models like exemplar based models or Categorization by Elimination.