| dc.description.abstract | Fairness is an inherently important aspect of decision-making in general, and with automated algorithms increasingly making decisions impacting our lives it is of paramount importance that the decisions they make be fair according to some fairness criterion and that this criterion is open to scrutiny. This is important both to ensure people are treated fairly, but also to ensure that the public, in general sceptical of machines having power, doesn’t stop the promising field of automated decision-making in it’s tracks. This thesis explores a Bayesian approach to optimizing decisions in settings where we have little data and high uncertainty. We seek to optimize the decisions with regard both to the utility of the decision-maker and to the decisions being fair. Finally we seek to take the decisions that give the optimal utility and fairness not only for the decisions taken at the present, but for all decisions taken now and in the future, by also placing a value on the information gained from our decisions. The results we were seeking to replicate did not, and our hypothesised method for improving that method resulted in no improvement. We therefore offer a framework for exploring approaches related to methods for optimizing a combination of fairness and utility, and of different utility functions, along with some reflections on the failures of the methods tested and possible other approaches to test out going forward. | eng |