Overall Objectives
Research Program
Application Domains
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, Yoann Isaac, Mostepha Redouane Khouadjia, Hélène Paugam-Moisy, Marc Schoenauer, Michèle Sebag.

This recently created 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.

Image understanding.

We continued our effort on the development of model-based image understanding approaches (e.g.  [71] ). In [18] , we have proposed a method for simultaneously segmenting and recognizing objects in images, based on a structural representation of the scene and on a constraint propagation method. Theoretical guaranties have been provided along with a quantitative assessment on healthy and pathological brain structures in magnetic resonance images. Within the ANR project LOGIMA (collaboration with ECP, Telecom PariTech and TU Dresden), our goal is to introduce a new lattice-based representation and reasoning framework suited for dealing with spatial objects in the presence of uncertainty. This framework associates under the aegis of general lattice theory ingredients from mathematical morphology, description logics and formal concept analysis. A first development of this framework can be found in [7] where it has been exploited for the definition of abductive reasoning services and applied to high-level image understanding. Several theoretical issues have been raised in the development of this new framework. Some of them were tackled in [25] , [24] , [30] . In [25] , we have shown how mathematical morphology operators can ben defined on general concept lattices. Explicit join-commuting and meet-commuting operators are defined either from particular valuations on the corresponding lattice or from the decomposition of their elements. In [24] , we extended our mathematical morphology based adductive reasoning to multivalued logics, hence allowing us to deal with several uncertainty and imprecision phenomena. In [30] , metrics between bipolar information - where the information is represented by a positive/preference part and a negative/constraint part - have been introduced based on particular dilations.

Structured learning.

With motivations in bio-informatics and brain computer interfaces, the goal is to take into account priors about the spatio-temporal structure of the underlying phenomenon in order to propose a generative model of the data.

In the context of Yoann Isaac's PhD (Digiteo Unsupervised Brain project), in collaboration with CEA LIST, the goal is to design a representation endowed with appropriate invariance properties. Specifically, within the framework of sparse dictionary coding, we have introduced new priors allowing us to capture both spatial and temporal regularity of multivariate brainwave signals [54] . The learning/optimization criterion, while being multivariate, contains several non-differentiable terms, raising new optimization issues; the proposed approach extends the classical split Bregman iterations algorithm to the multi-dimensional case with several non-differentiable terms [37] .

In the context of regulatory gene networks, the challenge is to combine probabilistic inference (does a gene regulate another one) with relational learning (the set of genes is organized in a network). Ensemble learning approaches have been used to cope with the imbalanced nature of the data, e.g., bagging Markov logic networks or boosting operator-valued kernel-based regressors [55] , [64] . Another issue, regarding the indeterminacy of the models due to the data sparsity, is addressed through prior-guided regularization beyond model sparsity such as orthogonality [8] or stability [16] .

In the domain of medical imaging, the exploitation of computational tomography data to model tumor physiology is hindered by the huge noise level; the multi-task setting is leveraged to provide a better robustness to noise [51] .

Robotic value systems.

Within the European SYMBRION IP, investigations on the preference-based reinforcement learning were continued in Riad Akrour's PhD, where the robot demonstrations are assessed by the expert and these assessments are used to learn a model of the expert's expectations. In [67] , this work had been extended and combined with active learning to yield state-of-the art performances with few binary feedbacks from the expert. The work has first concentrated this year on handling the noise due to expert's mistakes [53] , and bridging the reality gap when porting the algorithms on real robots (e-pucks and one Nao robot) – these results will be published in Riad Akrour's PhD dissertation, to be defended in Spring 2014.

Algorithm Selection as Collaborative Filtering.

The crucial issue when addressing algorithm selection problems is to be able to come up with features that can describe the problems: with representative features, algorithm selection amounts to supervised learning. However, except for some rare domaines (e.g., SAT, [73] ), no satisfying set of features exists. However, algorithm selection can also be viewed as a recommendation problem, and tackled by collaborative filtering: users more or less like movies, and similarly, instances like algorithms as much as these algorithms are efficient in solving it. Applying collaborative filtering leads to designing a latent feature space in which the representation of the problems is highly adapted to the algorithm selection problem. A critical issue in collaborative filtering is the 'cold start' problem, that is making recommendations for a brand new user/problem instance. This issue has been handled by a surrogate model of the latent factors, mapping the initial features onto the latent ones. The Algorithm Recommender System has been successfully applied to 3 different domains: Satisfiability, Constraint Programming, and Continuous Black-Box Optimization (data from the COCO platform, see Section 5.4 ) [59] .

Social Networks with insider information.

The analysis of social networks based on the contents and structure of information exchanges most often pertains to descriptive learning, e.g., explaining the growth of the communication graph or investigating the sensitivity of existing algorithms to hyper-parameters [31] . In the Modyrum context (coll. SME Augure), a supervised learning perspective is investigated, taking advantage of the fact that experts already know part of the sought results in some specific domains of interest. Based on e.g., Twitter and blogs data, the goal is to define generic features and supervised learning algorithms, enabling to characterize the targets of interest depending on the public relation focus.

Multi-objective AI Planning.

Within the ANR project DESCARWIN ( ), Mostepha-Redouane Kouadjia continued his work on the multi-objective approach to AI Planning using the Evolutionary Planner Divide-and-Evolve (DaE), that evolves a sequential decomposition of the problem at hand: each sub-problem is then solved in turn by some embedded classical planner [70] . Even though the embedded planner is single-objective, DaE can nevertheless handle multi-objective problems. Current work includes the implementation of the multi-objective version of DaE, the definition of some benchmark suite, and some first numerical experiments, comparing in particular the results of a full Pareto approach to those of the classical aggregation method. These works resulted in 3 conference papers recently accepted, introducing a tunable benchmark test suite [39] , demonstrating that the best quality measure for parameter tuning in this multi-objective framework is the hypervolume, even in the case of the aggregation approach [41] , and comparing the evolutionary multi-objective approach with the aggregation method, the only method known to the AI Planning community [38] . A sum-up of these recent results have been published at IJCAI [40] .