Using a Bayesian model for confidence to make decisions that consider epistemic regret

This paper is motivated by a desire to understand the role played by regret in high-stake decision making. The scenarios of interest are those which involve only sparse data being available at the time the decision needs to be made, but further data being made available subsequent to the decision, after which regret is experienced. Previous work has considered regret in scenarios where the decision is based on data that makes it ambiguous which of several hypothesis are true, but where the data provided after the decision removes all this ambiguity. This paper considers epistemic regret, ie regret in scenarios where the data provided after the decision does not remove all the ambiguity. This focus motivates the development of a Bayesian model for confidence. This model for confidence plays a key role in the development of a decision making model that considers epistemic regret. Results in the context of an accessible example demonstrate the ability of the model to predict how epistemic regret will result in intuitive, if sub-optimal, decision making. It is shown that increasing initial confidence reduces the extent to which sub-optimal decision making occurs, while increasing the cost associated with regret increases the extent to which it occurs.