Supervisors:

• Prof Sylvie Thiebaux (computer science)
• Dr Ignacio Ojea (philosophy)

The student will also have access to the expertise of other members of the Humanising Machine Intelligence Project (HMI), including Dr Pamela Robinson, Dr Claire Benn (philosophy), Dr Alban Grastien and Dr Hanna Kurniawati (computer science).

As autonomous agents become ubiquitous, it is increasingly important for these decision-making systems to align with our ethical values (Russell, Stuart, 2019.). One way to incorporate ethical values is to impose constraints on the decision problem, but there is so far no principled way of doing this. Furthermore, the incorporation of constraints can increase the complexity of the decision problem.

In this project, we will consider Markov Decision Processes (MDPs) and/or related stochastic shortest paths problems as the decision-making paradigm. Constraints on such decision problems can be placed on actions, states, state/action trajectories, or on the reward function. But the complexity of solving for the optimal policy depends on the form of the constraints, the formalism used to encode them, the type of policy sought (deterministic vs. stochastic), and on the solution method used for solving the MDP. Successfully implementing ethical constraints will require a careful analysis of the different options.

The literature on ethical constraints for autonomous decision making is still nascent (Bringsjord, Arkoudas, and Bello 2006, Lindner et al, 2019, Svegliato et al. 2020, Nashed et al. 2021). There are opportunities to make significant contributions to it, both to the project of rigorously formalising important philosophical concepts, and to the project of exploring the computational limitations, complexity and implementation of these formalisations.

In this thesis, we will first identify and taxonomise the class of ethical constraints that are both philosophically significant and formalisable as constraints over MDPs. Second, we will compare the levels of added complexity that those formalised constraints add to decision problems. Third, we will implement the most promising formalisations within an MDP planner.