Reinforcement Learning (RL) is a learning paradigm in which the agent learns from its environment through trial and error. Deep reinforcement learning (DRL) algorithms represent the agent’s policies using neural networks, making their decisions difficult to interpret. Explaining the behaviour of DRL agents is necessary to advance user trust, increase engagement, and facilitate integration with real-life tasks. Semifactual explanations aim to explain an outcome by providing “even if” scenarios, such as “even if the car were moving twice as slowly, it would still have to swerve to avoid crashing”. Semifactuals help users understand the effects of different factors on the outcome and support the optimisation of resources. While extensively studied in psychology and even utilised in supervised learning, semifactuals have not been used to explain the decisions of RL systems. In this work, we develop a first approach to generating semifactual explanations for RL agents. We start by defining five properties of desirable semifactual explanations in RL and then introducing SGRL-Rewind and SGRL-Advance, the first algorithms for generating semifactual explanations in RL. We evaluate the algorithms in two standard RL environments and find that they generate semifactuals that are easier to reach, represent the agent’s policy better, and are more diverse compared to baselines. Lastly, we conduct and analyse a user study to assess the participant’s perception of semifactual explanations of the agent’s actions.

Intelligent Personal Assistants (IPAs) have become increasingly ubiquitous, yet they remain primarily reactive, non-personalised, and inscrutable. Moreover, concerns regarding user control, data stewardship, and communication design persist in the literature. Aiming to shape an appropriate human-assistant interaction framework, we organised a multidisciplinary expert discussion focusing on proactive IPAs for time management – a bounded yet complex domain that may catalyse identifying and tackling paradigmatic challenges. We invited the experts to propose, debate, and chart interaction scenarios, desired characteristics and constraints, and user modelling for proactive IPAs. This paper presents the thematic analysis of the discussion and the resulting interaction diagram. The ability of the user to scrutinise the assistant’s models that underpin personalisation and to receive adequate explanations were identified to be of paramount significance. Moreover, a proposed onboarding and control mechanism that may help align the user’s perception of the system and the system’s actual capabilities is discussed.

Intelligent personal assistants (IPAs) have become widely available, yet they remain primarily used for discrete, straightforward tasks. By contrast, both user studies and literature reviews indicate that IPAs of the future are to be personalised, proactive, and capable of performing elaborate undertakings. Such systems would have to be based on complex and dynamic user and context models. We believe that scrutability – i.e. the ability of the user to actively study and modify the models towards tuning personalisation – could emerge as an essential element of such a human-assistant interaction paradigm. Yet, to the best of our knowledge, no work so far has investigated how the principles of scrutability, as presented in [21], relate to the context and novel challenges raised by the proactive IPAs and how scrutability could facilitate effort-efficient control of the assistants. This paper introduces our vision of the confluence of the research fields of IPAs and scrutability, presents a diagram of the proposed interaction structure, and reanalyses data from user studies originally presented in [11, 39] to better understand user expectations regarding scrutability and proactivity of IPAs.