Adaptive Elicitation of Rank-dependent Utilities
------
Stage rémunéré dans le cadre du projet Elicit soutenu par Idex Sorbonne Universités.
------
Research internships in Paris, at Lip6 (www.lip6.fr), the Computer Science laboratory of the University Pierre and Marie Curie (UPMC) for preparing the final Master project (6 months, with approximately 500 euros of monthly stipend) are available. French language not required.
* équipe Décision, laboratoire LIP6, CNRS - UPMC
* Supervisors: Patrice Perny, Christophe Gonzales, Paolo Viappiani and Angelina Vidali.
* 6 months paid internship (approximately 500 euro/month).
Description: Preference elicitation is a key problem for decision support systems and Artificial Intelligence in general. AI tools need to make decisions on behalf of a user (or decision maker), but in order to be effective, these systems need a model for the preferences of the user. Preferences are however only partially known and expensive (from a computational or cognitive point of view) to acquire; therefore preference elicitation constitute a veritable "bottleneck".
A number of recent works have proposed solutions to this problem that are adaptive in the sense that a belief (either probabilistic or based on strict uncertainty) is maintained over the possible instantiation of utility parameters; questions or queries are asked in order to obtain the most informative preference informations.
Almost all of these approaches assume models that are linear in the parameter space (as in additive utility models). However these models fail to represent certain common decision patterns that are well explained by descriptive decision theory, more precisely by rank-dependent utility (RDU) and cumulative prospect theory (CPT).
These models can capture, among other things, the user's attitude towards risk.
In this context, we are also interested in eliciting preferences in multi-attribute settings, where local utilities are aggregated into a single utility score using the Choquet integral.
The goal of the internship is to develop an algorithmic framework for the elicitation of complex preference models (according to RDU and CPT) and to test their effectiveness in simulations.
The internship will take place in the framework of a project financed by the IDEX Sorbonne University, involving our partners at the INSEAD business school.
Ideal profile: the candidate should have expertise in Operation Research (in particular linear and integer programming) and Artificial Intelligence. The candidate will have an interest in decision theory, multi-criteria optimization and algorithmics. It is expected that the student will have to develop simulations using a solver like CPLEX or Gurobi.
Contacts: paolo.viappiani@lip6.fr, patrice.perny@lip6.fr.
--------------------------------------------------------