Phd Position F/m

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat **:CDD

**Niveau de diplôme exigé **:Bac + 5 ou équivalent

**Fonction **:Doctorant

**A propos du centre ou de la direction fonctionnelle**:
The Inria centre at Université Côte d'Azur includes 37 research teams and 8 support services. The centre's staff (about 500 people) is made up of scientists of different nationalities, engineers, technicians and administrative staff. The teams are mainly located on the university campuses of Sophia Antipolis and Nice as well as Montpellier, in close collaboration with research and higher education laboratories and establishments (Université Côte d'Azur, CNRS, INRAE, INSERM...), but also with the regiona economic players.
With a presence in the fields of computational neuroscience and biology, data science and modeling, software engineering and certification, as well as collaborative robotics, the Inria Centre at Université Côte d'Azur is a major player in terms of scientific excellence through its results and collaborations at both European and international levels.

**Contexte et atouts du poste**:
An added perk of this position is the idyllic location. Sophia Antipolis, where the INRIA center is situated, is renowned for its concentration of research centers, creating a vibrant atmosphere for intellectual exchange. Moreover, Sophia Antipolis is a short distance from the French Riviera, with the beautiful coastal towns of Antibes and Nice offering a high standard of living, stunning beaches, and a rich cultural scene.

**Mission confiée**:
**Research context**

To circumvent this issue, a lot of sub-optimal algorithms have been proposed so that good solutions can be found even if they are not the best solutions. Such approximation rely heavily on heuristics, which are typically hand-made and refined over the years by researchers. Due to their essence, those heuristics are often problem-specific and sub-optimal.

As Bengio et al. [5] phrase it, the “techniques, and the parameters controlling them, have been collectively learned by the community”. This is why a new approach has been proposed recently: using neural networks to learn the heuristics. They are trained on a large set of instances of the problem, and learn to generalize to unseen instances. Those new heuristics can be trained within days and be specialized onto the specific distribution of problems seen during training. Moreover, they require little human intervention and profit from the research done in the field of machine learning. This is a promising way for enterprises to quickly build a solver for their very specific problem, where available solvers may not be adapted.

The idea to train a model incorporated inside a solver is not new [24], but recent advances in deep learning made it possible to train a model to solve a problem end-to-end without relying on hand-made features nor external algorithm [28, 19, 10]. Those methods are encapsulated in the Neural Combinatorial Optimization (NCO) [4, 6] framework. Such methods showed good results when compared to classical solvers on small instances. But their performance quickly degrades when the problem size increases, and they do not generalize well to big instances [18, 6]. This is why currently the best approach is to use the NCO model within another optimization algorithm such as the Monte Carlo Tree Search [13] or as a policy space search [7]. This causes a much longer computation time and it raises the question of the generalization capabilities of the NCO models.

Interestingly, a field of research called “Adaptive Neural Computation” [16] has been developed with the goal of improving the generalization capabilities of neural networks.

Specifically, the idea of an “Adaptive Computation Time” (ACT) is that the network should adapt its computational depth to the complexity of the input [2, 12]. It lets the model “ponder” deeper for hard tasks, based on what it has learned during training. Researchers developed several methods to achieve this goal and showed that it allows the network to be trained on easy examples and still generalize well to hard examples [3]. Those methods have not been applied to NCO models and it remains to be seen if they can improve the generalization capabilities of NCO models. This line of work follows the algorithmic alignment hypothesis from Xu et al. [29].

**Research activities**:
The goal of this project is to study the generalization capabilities of NCO models and to propose new methods based on the theory of adaptive neural computation. Such work aligns well with the expertise of the COATI team, as it involves graphs, optimization, and deep learning - all areas the team has experience in. Their past work includes research on plain graph problems [27, 15, 20, 21], theoretical deep learning [8, 9, 11], and optimization problems using reinforcement learning [14, 23].

**Improve Algorithmic Alignment of GNNs using Methods fro