Resource-dependent rationality: the brain as a Bayesian sampler

This project aims to examine human decision making when it is optimal and when it fails. Specifically, it aims to test the possibility that the way the brain makes inferences about the world can be approximated though the analogy of a Bayesian sampler. This means that humans are in principle able to optimally reason probabilistically, but reliably fail to do so as soon as necessary information or cognitive resources are lacking.
Despite a vast amount of literature on decision-making and rationality, there is a notable scarcity of plausible models that can reconcile accounts of human behaviour in both "small" and "large" worlds (in terms of Savage, 1954). Such a model would need to explain why people behave within the predictions of rational choice theory so well that one can witness a tremendous success of Bayesian modelling approaches (Griffiths, Sanborn, Canini & Navarro, 2008; Oaksford & Chater, 2007), while systematic deviations from this behaviour occur reliably in real-world decision-making scenarios. One such possibility is that the brain is a Bayesian sampler (Sanborn & Chater, 2016). According to this explanation, people do not calculate or even represent probabilities. Instead, they understand the world by sampling from the evidence available to them. This sampling process can be approximated by Bayesian models, such as Markov Chain Monte Carlo (MCMC) algorithms, which only take into account current states, and not prior probabilities, when transitioning between sampling. In particular, Sanborn & Chater (2016) show that such a Bayesian sampler makes classic reasoning errors documented in the Heuristic and Biases literature (Tversky & Kahneman, 1974), unless an infinite number of samples can be taken. Such a sampler would have to navigate a landscape of probabilities. If, for instance, the sampler is drawn to a particular location in the search space, subsequent samples may tend to be nearby, reproducing the well-documented anchoring effect (e.g. Ariely, Loewenstein & Prelec, 2003). Thus, this approach allows for precise predictions of computational limitations, as well as modelling of human behaviour when it is within the broader boundaries of rationality, and when it is necessarily not.[...]
Much has been written and researched in the realm of rationality and decision-making. The proposed project offers to extend the literature in the sense of testing a plausible model of when and why people are able to make rational choices, and when they are not. Crucially, this will take into account both the availability of information and cognitive resources. This is useful, because it allows for more precise predictions than current models, which make imprecise and inflexible predictions, and do not sufficiently take these factors into account.