Published on December 13, 2018 | Updated on January 29, 2019

CORTEX conference by Philipp Schwartenbeck

September 17th, 2015

Decision making as active Bayesian inference: Do we minimise surprise or maximise utility in choice behaviour?

Recently, a computational framework of choice behaviour has been proposed that casts decision-making as an inferential process based on variational message passing in the brain. Here, planning and decision-making are understood as approximate Bayesian inference, where agents minimise surprise about future states based on a (generative) model of the environment. 
One prediction from that perspective is that beliefs about choice have a precision, which can be understood as confidence in policy selection. Precision has a similar computational role as an inverse temperature in classical softmax choice rules but, crucially, is not a free parameter but has a Bayes optimal solution that has to be inferred. I will illustrate this inference process in an experimental paradigm and briefly discuss the hypothesis that (expected) precision is encoded by neuromodulators, in particular dopamine. 
Another important prediction is that we try to minimise surprise when making decisions, which reduces to KL control if there is no uncertainty about the hidden states that cause outcomes. Here, the idea is that agents try to minimise the difference (KL divergence) between preferred and likely outcomes. Importantly, this implies that agents try to maximise utility as well as the entropy over outcomes, i.e. ‘keep their options open’, as opposed to maximising utility alone. I will present a study to test this assumption and preliminary results, as well as future directions of our work.