ORUS

Reliability, Optimization and Uncertainties

Principal investigator: Sébastien Destercke (Heudiasyc)

 

 

The ORUS project aimed at building bridges between the fields of optimization, reliability analysis and uncertainty. This project has benefited from a rich environment, building upon different funding sources such as the Labex, the ANR agency (project RECIF) or Chinese CSC thesis grants. This explains why it could thrive, despite the leave of Michael Poss, initially in charge of the optimization part. 

 

 

 

Two of the main identified goals of ORUS were, on the one hand, the design of optimal systems under severe uncertainty (an important task in early-stage design) and, on the other hand, the optimal elicitation of expert opinion to reduce our uncertainty about the system reliability, taking for instance inspiration from recent works in preference learning.

 

  • Regarding the design of optimal systems, M. Sallak and S. Destercke are currently supervising Lanting Yu’s PhD, whose work [Lanting Yu, IPMU 2016] focuses on comparing systems when failure probabilities are ill-known. Lanting Yu’s internship was funded by the Labex. Her work complements the initial work of M. Sallak, M. Poss and S. Destercke published at ICRAM 2014 and concerning the estimation of reliability under unknown dependencies.
  • Regarding the optimal elicitation of expert opinions, the work performed by N. Ben-Abdallah during her post-doctoral position within the ANR project RECIF allowed us to develop computationally efficient myopic techniques, where we only search the next optimal question, at each iteration, to query experts. These techniques, which have been published in a top AI conference [Ben Abdallah, UAI 2015] and in optimization journals [Ben Abdallah, EJOR], are currently being experimentally tested among experts of our network in Valenciennes.  This work is currently pursued by the internship (funded by the Labex) of Jospeh Rocca, who investigates the much harder sequential problem, where a whole sequence of optimal questions is sought instead of a single question that is optimal for the next step.

 

We currently plan to pursue in parallel these two directions, building upon the results we already have, and trying to involve even more real experts in order to test the empirical validity of our approaches. In the future, two lines of possible research have been identified: the involvement of a researcher in optimization (this seems possible, with the arrival of F. D'Andreagiovanni in Heudiasyc Laboratory), or the involvement of researchers in social science and cognitive psychology in practice (which can be done through the Sorbonne Université Network).