Section: Scientific Foundations

Designing criteria

Participants : Jamal Atif, Nicolas Bredèche, Cyril Furtlehner, Yoann Isaac, Victorin Martin, Jean-Marc Montanier, Hélène Paugam-Moisy, Marc Schoenauer, Michèle Sebag.

This new 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.

In the context of the ANR project TRAVESTI (http://travesti.gforge.inria.fr ) which ended this year, we have worked to address more specifically the inverse pairwise Markov random field (MRF) model. On one side we have formalized how the Ising model [64] can be used to perform travel time inference in this context (submitted to AISTATS). Extending the ordinary linear response theory in the vicinity of the so called “Bethe reference point” instead of the interaction-free reference point [67] , we were able to provide explicit and tractable formulas to the Plefka's expansion and the related natural gradient at that point. These can be used into various possible algorithms for generating approximate solutions to the inverse Ising problem which we do investigate now. In parallel we have proposed also in [67] , a method based on the “iterative proportional scaling” to learn both the factor graph and the couplings by selecting links one by one (submitted to AISTATS). In the Gaussian MRF case this can be implemented efficiently due to local transformations of the precision matrix after adding one link, which is unfortunately not possible in the Ising MRF. It is competitive with $L_0$ based approach in terms of precision and computational cost, while incidentally the $L_1$ based method potentially cheaper is not working well for this problem. The flexibility of the method offers in addition the possibility to combine it with spectral constraints like walk-summability with belief propagation or/and graph structure constrain to enforce compatibility with BP. With no additional computational cost we get a complete set of good trade-offs between likelihood and compatibility with BP. Concerning the analysis of the belief propagation we establish in [60] some sufficient condition for encoding a set of local marginals into a stable belief propagation fixed point. Following some work of last year concerning the modeling of congestion at the microscopic level we have finalized in [9] the analysis of a new family of queuing processes where the service rate is coupled stochastically to the number of clients leading a large deviation formulation of the fundamental diagram of traffic flow.

Multi-objective AI Planning.

Within the ANR project DESCARWIN (http://descarwin.lri.fr ), Mostepha-Redouane Kouadjia worked on the multi-objective approach to AI Planning using the Evolutionary Planner Divide-and-Evolve, that evolves a sequential decomposition of the problem at hand: each sub-problem is then solved in turn by some embedded classical planner [72] . 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 [45] , 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 [46] , and comparing the evolutionary multi-objective approach with the aggregation method, the only method known to the AI Planning community [44] .

The parameter-tuning algorithm designed for DAE, called Learn-and-Optimize, was published in the selection of papers from Evolution Artificielle conference [26] . Though originally designed for Evolutionary AI Planning, the method is applicable to domains where instances sharing similar characteristics w.r.t parameter tuning can be grouped in domains.

Image understanding.

Sequential image understanding refers to the decision making paradigm where objects in an image are successively segmented/recognized following a predefined strategy. Such an approach generally raises some issues about the “best" segmentation sequence to follow and/or how to avoid error propagation. Within the new sequential recognition framework proposed in [8] , these issues are addressed as the objects to segment/recognize are represented by a model describing the spatial relations between objects. The process is guided by a criterion derived from visual attention, specifically a saliency map, used to optimize the segmentation sequence. Spatial knowledge is also used to ensure the consistency of the results and to allow backtracking on the segmentation order if needed. The proposed approach was applied for the segmentation of internal brain structures in magnetic resonance images. The results show the relevance of the optimization criteria and the relevance of the backtracking procedure to guarantee good and consistent results. In [70] we propose a method for simultaneously segmenting and recognizing objects in images, based on a structural representation of the scene and on a constraint propagation method. Within the ANR project LOGIMA, our goal is to address sequential object recognition as an abduction process [69] .

Similar principles are at the core of Yoann Isaac's PhD (Digiteo Unsupervised Brain project), in collaboration with CEA LIST. The dictionary-learning approach used to decompose the EEG signal is required to comply with the structure of the data (e.g. spatio-temporal continuity; submitted).

Robotic value systems.

Within the European SYMBRION IP, a key milestone toward autonomous cognitive agents has been to provide robots with internal or external rewards, yielding an interesting or competent behavior. Firstly, an objective-free setting referred to as open-ended evolution has been investigated [17] , [5] , where the criterion to be optimized is left implicit in the reproduction process. in the Secondly, preference-based reinforcement learning has been investigated 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 [21] , this work has been extended and combined with active learning to yield state-of-the art performances with few binary feedbacks from the expert. The hormone-based neural net controller first proposed by T. Schmickl et al. has been thoroughly analyzed and simplified in collaboration with Artificial Life Laboratory from Graz [34] .