Overall Objectives
Research Program
Application Domains
New Software and Platforms
New Results
Bilateral Contracts and Grants with Industry
Partnerships and Cooperations
XML PDF e-pub
PDF e-Pub

Section: New Results

Designing criteria

Participants : Jamal Atif, Aurélien Decelle, Cyril Furtlehner, Yoann Isaac, Alexandre Quemy, Yann Ollivier, Marc Schoenauer, Michèle Sebag.

This SIG, rooted on the claim that What matters is the criterion, aims at defining new learning or optimization objectives reflecting fundamental properties of the model, the problem or the expert prior knowledge.

A statistical physics perspective

With motivating applications in large scale inference problems like traffic congestions we are pursuing our quest of practical solutions to inverse problems like in [39] where a method is proposed to invert a Gaussian Markov random field with topological and spectral constraints well suited to subsequent use of belief propagation as inference algorithm (see for the implementation). A more specific model for traffic inference has also been developped in [11] . A method adapted to the generalized belief propagation framework, aiming at adressing directly and systematically the loop corrections without loss of scalability is about to be completed.

Multi-objective ATC

The new Bayesian approach of Air Traffic Control belongs to this SIG, but was described in the Section 4.2 . Main publications are Gaétan Marceau's PhD [4] and the corresponding PPSN paper [38] , [59] .

Programming by Feedback

Riad Akrour's PhD work on Preference Based Learning [1] culminated with the addition of a model for the user's competence in the interactive learning loop. In the resulting original paradigm, the user is sequentially proposed a series of behaviors and is only asked "Hot-or-cold" questions. The Programming by Feedback paradigm [15] will hopefully initiate a general way to allow non-digitally-proficient users to nevertheless control the behavior of software-based agents in their environment.

Multi-objective AI Planning

This activity had almost stopped since the end of the DESCARWIN ANR project. However, a productive intership resulted in some new benchmarks in the ZenoTravel domain together with an exact solver ensuring the knowledge of the true Pareto front [41] , [40] .

Algorithm Selection

Algorithm Selection can be viewed as a Collaborative Filtering problem, in which a problem "likes" an algorithm that is able to solve it. Initiated during Mustafa Misir's ERCIM postdoc in 2013, this idea has also been applied for Process Management [43] , and is the basis of François Gonards's PhD funded by IRT SystemX in the context of aeronautics and car industry.

Outlier rejection in classification

An original approach based on One-Class SVM has been proposed during Blaise Hanczar's on year delegation at TAO [28] .

Learning sparse representations by auto-encoders

Auto-encoders (AE) are a widely used tool for unsupervised learning, which consists of a neural network trained to reconstruct its own input via smaller-dimensional layers. The usual training criterion is the reconstruction error, however, the usual justification for AE is to learn a more compact data representation. In [62] we formalize this latter criterion using Minimum Description Length (MDL) and establish a comparison with the traditional reconstruction criterion. The MDL criterion has an interpretation as a denoising reconstruction and fully determines an optimal noise level, contrary to the literature on denoising AEs. More surprisingly, AE (aka Auto-associators) can also be used to learn sparse representations in the context of supervised learning [51] .