Abstracts (ordered alphabetically by speaker's last name)

Martin Mose Bentzen

The principle of double effect applied to thought experiments in ethics – a formal approach

The principle of double effect states conditions for ethically acceptable behavior when there are both positive and negative consequences (effects) at the intended outcome of an act. The act itself must be good or neutral, the negative consequence must not be intended whereas the positive must be, the negative consequence must not be a means to obtain the positive consequence, and the positive consequence must be proportionally preferable to the negative consequence. In this talk, I propose extensions of well-known formal models of consequentialist reasoning from game theory and stit theory. In particular, I suggest how to handle intentions, means-ends reasoning and proportionality of several positive or negative aspects of an event. I apply the formal methods in an analysis of a couple of thought experiments from ethics. I then discuss limitations of the models and of the principle of double effect.

This research is part of an ongoing project of providing a foundation for machine ethics through formal semantics and logic.

Jens Ulrik Hansen

Reasoning about influence in social networks

In this talk, I will discuss a recent body of work, within logic, that aims at modeling how beliefs (or opinions or preferences) may change in networks of agents that are prone to social influence. The work was initiated by Patrick Girard, Fenrong Liu, and Jeremy Seligman that proposed a simple model of “belief change in the community” and studied the long-term dynamics of this model. This model was later expanded by Zoé Christoff and the speaker to allow for more complex social phenomena such as pluralistic ignorance. Still, this model assumes that the agents themselves do not reflect on the fact that their peers are prone to social influence, as well. To remedy this lack of social intelligence, the model has recently been further extended by Zoé Christoff, Carlo Proietti, and the speaker, by introducing knowledge and learning into the picture.

Ron Petrick

Knowledge-level planning for task-oriented social interaction

An intelligent agent coexisting with humans must not only be able to perform physical tasks, but must also be able to interact with humans in a socially appropriate manner. In many settings, this involves the use of social signals like gaze, facial expression, and language. To take full advantage of such information, the results of social signals should be represented as part of the agent's domain model, and utilised by its reasoning and decision making processes, for instance when selecting actions to perform in the world. Using the example of a robot in a simple bartending scenario, I will describe an application of knowledge-level planning to the problem of task-based social interation, and discuss how knowledge, action, and social information are modelled within the PKS (Planning with Knowledge and Sensing) planner as an instance of planning with incomplete information and sensing. I will also discuss current work that seeks to extend this approach to reason about multiagent knowledge, and to encode certain types of social protocols.

Rasmus Rendsvig

Autonomous "Meta-Models" for Dynamic Epistemic Logic

The typical modus operandi applied by modelers working with Dynamic Epistemic Logic is to manually pick an update judged to be suitable based on the current state of the system. Hence the system level model is seldomly complete, lacking information about update choices in runs counterfactual to the chosen initial state.

In this talk, we will introduce a framework for formally defining sequential DEL models in a local, DEL-like fashion, useful in producing complete models. We call the introduced "meta-models" DEL machines. A DEL machine contains a set of rules by which the current epistemic state model gives rise to a choice of the next update to be applied. The rules are akin to knowledge-based programs and one-step planning problems. DEL machines work autonomously, picking the next update without the need for the modeler's intervention.

Once introduced and exemplified, we will compare DEL machines to an alternative approach to meta-model construction, namely DEL protocols. One aspect is heuristic, where we argue that DEL machines are less cumbersome to use in model construction. Secondly, the two approaches will be formally compared via representation theorems, linking the two frameworks by comparing the classes of epistemic forests they can generate.

The talk based on joined work with Suzanne van Wijk and Alexandru Baltag.

Sonja Smets

A logical analysis of belief formation and doxastic influence based on evidence and trust

I use the tools of dynamic epistemic logic to reason about the process of belief change triggered by evidence and trust in a social network. We model the epistemic and doxastic states of agents, the strength of the agent's own private evidence, the strength of their friends' evidence, as well as the trust that the agents have in their friends (and in themselves) as a reliable source of information. Our formal approach brings together two ingredients: 1) the setting of 'justification models', a formal qualitative representation of an agents' beliefs, evidence and justification [1] and 2) the work on belief dynamics and doxastic influence in social networks [2]. In this talk we use results from dynamical systems to characterize the conditions of doxastic stability in the social network in the long run. This presentation is based on joint work with A. Baltag and F. Liu.

[1] Virginie Fiutek, Playing with Knowledge and Belief, PhD thesis, ILLC, 2013.

[2] Fenrong Liu, Jeremy Seligman, and Patrick Girard: Logical Dynamics of Belief Change in the Community, Synthese, Volume 191, Issue 11, pp 2403-2431, 2014.