The main objective of the project is to propose a novel approach of distributed, scalable, dynamic and energy efficient algorithms for managing resources in a mobile network. This new approach relies on the design of an orchestration mechanism of a portfolio of algorithms. The ultimate goal of the proposed mechanism is to enhance the user experience, while at the same time to better utilize the operator resources. User mobility and new services are key elements to take into account if the operator wants to improve the user quality of experience. Future autonomous network management and control algorithms will thus have to deal with a real-time dynamicity due to user mobility and to traffic variations resulting from various usages. To achieve this goal, we focus on two central aspects of mobile networks and intend to design distributed learning mechanisms in non-stationary environments, as well as an orchestration mechanism that applies the best algorithms depending on the situation. The first main aspect to be addressed is the management of radio resources at the RAN (Radio Access Network) level. In current (LTE) and future (LTE-A) cellular systems, interference appears as a bottleneck for providing high data rates and seamless connectivity to the end-user. To reduce interference it is possible either to coordinate the transmissions of neighboring base stations (BS) so as to avoid simultaneous transmissions on the same radio resources or to allow BSs to cooperate: two or more BSs combine their transmissions towards a single user in order to increase its data rate. Both cases require distributed learning algorithms. The second aspect is the management of the popular contents users want to get access to. In a Content Delivery Network (CDN), popular content is disseminated and stored in cache servers as close as possible to the demand to avoid delay in access. How to place servers in the network and replicated contents in the servers are traditional issues in CDNs. In mobile CDNs, things are exacerbated because of the changing and unpredictable environment characterized by spatial and temporal changes in the traffic demand, user mobility and variable channel conditions. The way the project intends to tackle these problems is based on a “learn to learn approach”. If we think about BSs and cache servers as autonomous entities seeking to optimize a global objective function and able to take decisions based on incomplete information, the notion of distributed learning arises naturally. There are numerous approaches along these lines and each mechanism has its own characteristics in terms of needed information, type of achieved equilibrium, convergence speed, and stability. Each mechanism can also be tuned thanks via an array of parameters. The problem is exacerbated in non-stationary situations due to mobility, traffic demand or radio channel variations. The originality of the project thus lies in its objective of building a portfolio of distributed learning algorithms that are then to be orchestrated. To account for learning in the presence of non-stationary processes, we intend to use the theory of stochastic approximation in order to develop robust versions of existing learning schemes. Orchestrating a portfolio of learning algorithms is, in many regards, similar to the literature on “learning with expert advice”, so our goal will be to devise adaptive learning schemes that select dynamically between different learning schemes so that their long-term learning power exceeds the regret of any individual “expert”. Bringing together experts form both network and learning, NETLEARN ultimately intends to propose architecture and protocol adaptations for implementing our resource management algorithms.

The last decade has acknowledged the emergence of a new paradigm for building large-scale distributed systems: the microservice architecture. As opposed to the classical tightly coupled monolith architectures, this paradigm promotes small loosely coupled autonomous services, so called microservices. Such software services are to be small to encapsulate a well-defined functionality and thus facilitate bug isolation and maintenance. They should be loosely coupled to support independent software updates. Finally, they need to be autonomous to enable independent delivery to production. Adopting microservices becomes a general trend and experts predict that "by 2022, 90% of all new apps will feature microservices architectures". Among companies working around microservices are infrastructure providers, including GAFAM and EOLAS, major IT players, including Netflix and Uber, but also telco operators, including Orange. They all build on the container technology to facilitate functionality encapsulation, dependency isolation and standalone deployment, as required by microservices. However, the current used orchestration solutions (e.g., Kubernetes) struggle with the scale and the heterogeneity of microservices and fail to fix the performance problems fast enough. The challenge addressed by the SCALER project is to optimize the scaling of microservice-based networked services while satisfying their stringent IT and telco requirements. The proposal targets three research aspects, namely automatic microservice characterization, identification of microservice interaction patterns and microservice smart-scaling. SCALER brings together two academical and two industrial partners, namely LIG ERODS, INRIA Spirals, Orange Innovation and EOLAS B&D.

Human conversations pose significant challenges for automated processing due to their multimodal and dynamic nature. The MINERAL project aims to advance the comprehension of conversational situations by addressing key issues: leveraging non-verbal cues, incorporating semantic grounding, and understanding the dynamic discursive structure. Positioned at the intersection of textual, speech and visual modalities, the MINERAL project builds upon the extensive history of work on conversation recognition while focusing on the largely unexplored domain of multimodal conversational analysis. Leveraging datasets including TV fiction (Bazinga, REPERE), multiparty meetings (AMI, ELITR minuting corpus), and anonymized call center recordings from Orange, alongside the ChiCo dataset focusing on children's socio-cognitive development, the first objective of the MINERAL project will be to develop models that can perform multimodal contextual semantic, discursive, and pragmatic analyses of multi-party human-spoken conversations. To address the conversation in its globality, we will start by characterizing communicative acts, i.e. the smallest communication units within the conversation, that can be verbal or non verbal, then study the communicative structure, i.e. the relations between communicative acts. To assess the effectiveness of these models, the second objective of the project will be the automatic generation of a script (similarl to movie script) describing the content of a conversation considering semantic, discursive and pragmatic dimensions. We plan to assess the generalization capacity of our models across different use cases, industrial (meeting summarization) or academic, through research on children’s socio-cognitive development. A rigorous evaluation framework will be devised to measure the quality of generated content.