Machine Ethics is the study of how ethical reasoning can be computationally implemented. Approaches to Machine Ethics are divided into top-down and bottom-up approaches. In top-down approaches the machine reasons explicitly about ethics using utilities or symbolic rules while in bottom up approaches the machine learns ethical behaviour from observation. This project is primarily concerned with top-down approaches. Top down approaches to Machine Ethics generally draw upon theories of ethics from Philosophy such as utilitarianism or Deontic Logic. In general these philosophical theories deal with certainties, particularly when reasoning about the outcomes of actions - so they assume that some circumstance, such as a person being hurt, either will or will not be the outcome of the action. In some cases computer science has adapted these theories to incorporate uncertainty - for instance Markov Decision processes can be seen as a variation on utilitarianism that uses probabilities to calculated expected, rather than actual, utilities. However, in general, the effect of uncertainty on ethical reasoning has not been well-studied in computer science, and there is a lack of proposals for implementational frameworks that incorporate uncertainty into ethical reasoning.
This PhD would involve identifying key top-down approaches to the implementation of machine ethics and proposing mechanisms to adapt these approaches to account for uncertainty. Cases studies, preferably ones that could illuminate the difference between different implementational approaches, and between certainty and uncertainty would need to be developed in order to evaluate the proposals. Machine learning techniques could potentially be adopted in order to learn probabilities for the outcomes of actions that could then be incorporated into the ethical reasoning.
Minimum Viable PhD: The first year of the project would involve reading about top-down approaches to machine ethics and selecting at least two for further study. These could (though need not) be approaches already investigated by the primary supervisor and implemented within the Agent Infrastructure Layer that she has developed. In the first year the student would also consider the issue of uncertainty and the ways the selected approaches might be adapted to account for uncertainty. An initial proposal for a case study for evaluation of the project would be sketched out. Ideally this work would result in a position paper in one of a number of workshops devoted to machine ethics and ethics and AI. The second year would focus on an iterative process of implementing, evaluating and refining the proposed adaptations of the selected approaches to account for uncertainty. The third year would focus on writing up the results of the second year, both as papers in significant conferences and as part of the thesis. Implementations would be made publicly available (e.g., on GitHub).