Invited talk: Finding preferred relaxations of user preferences
Konstantin Schekotihin
Alpen-Adria-Universität Klagenfurt, AustriaAbstract:
Finding a product or a service that satisfies ones needs is a hard problem. In many cases the users of a recommender system are not experts in the field and do not know all features of the available items. In practical situations this leads to the formulation of a set of preferences that cannot be satisfied at once and the system does not return any results.
To resolve the problem multiple preference relaxation methods were suggested in the literature. However, it is not always clear which of the relaxations will be preferred by a user. In this talk we present an interactive approach aiming at identification of the preferred relaxation by automatically generating and asking a user a number of questions: whether the preferred item must have some feature or not. The answers to these questions allow a recommender system to find out the required relaxation.
Invited talk: Personalising Sequential Decision Processes under Uncertainty
Bruno Zanuttini
University of Caen, FranceAbstract:
Sequential Decision Making is one of the core problems in Artificial Intelligence. In general, it consists of taking decisions so as to maximise a long-term goal or reward. When the environment or outcomes of actions are uncertain, frameworks of choice are those of conformant and contingent planning, and of Markov Decision Processes, for which the basic problem is to find an optimal policy given a model of the task at hand.
Nevertheless, in many realistic applications, part of this model is not known to the decision maker. This talk will be about cases when this unknown part comes from the user, for instance because she is not able to formalise the rewards as numbers, or because she has clues about the environment different from those available to the agent. The talk will review some work tackling these issues, and present motivating applications as well as new, ongoing work so as to illustrate the relevant problems and possible approaches in the domain.
This talk will present work joint with Esther Nicart, Paul Weng, Hugo Gilbert, and Guillaume Desquesnes.
