Randomness perception: representativeness or encoding?

The probabilistic analysis of cognition is a recent framework that employs Bayesian statistics to model various aspects of the human cognitive system. After a brief description of this perspective, we present a Bayesian model of randomness perception (Griffiths and Tenenbaum, 2003, 2004). The randomness perception task is addressed in terms of the statistical problem of model selection: given a string, inferring whether the process that generated it was random or regular. A basic finding is that people rate sequences with an excess of alternation as more random than prescribed by information theory (overalternating bias). There are two explanations: local representativeness (Kahneman and Tversky, 1972) and the implicit encoding hypothesis (Falk and Konold, 1997). The measure random(X) of the Bayesian model was used in order to compare predictions derived from the explanations in a series of reaction times experiments. Results are discussed in relation to relevant methodological issues and future research.