Events in December 2023–January 2024
-
PhD defense Louis Sébastien Rebuffi: Algorithmes d’apprentissage par renforcement pour le contrôle de systèmes de files d’attente
PhD defense Louis Sébastien Rebuffi: Algorithmes d’apprentissage par renforcement pour le contrôle de systèmes de files d’attente
–
December 11, 2023Bien que l’apprentissage par renforcement ait été récemment principalement étudié dans le cas générique des processus de décisions markoviens, le cas des systèmes de files d’attente se dinstigue particulièrement. Pour compenser la taille de l’espace
d’état qui peut être extrêmement grande a priori, les algorithmes d’apprentissage doivent tenir compte de la structure des systèmes afin d’en extraire le plus d’information et de choisir le meilleur contrôle qui optimisent au mieux les performances
du système sur le long terme. Dans cette thèse, nous présentons des algorithmes construits à partir d’algorithmes classiques, adaptés au contexte des système de file d’attente, et nous étudions les performances de ceux-ci pour montrer une dépendance faible à l’espace d’états comparativement aux résultats obtenus dans le cas général.Jury:- Matthieu JONCKHEERE (DR CNRS, LAAS, Rapporteur)
- Ger KOOLE (professeur, Vrije Universiteil Amsterdam, Rapporteur)
- Nadia BRAUNER (professeure Université Grenoble Alpes, présidente)
- Bruno GAUJAL (DR INRIA, LIG, Directeur de thèse)
- Alain JEAN-MARIE (DR INRIA, Sophia Antipolis, Examinateur)
Thèse supervisée par Jonatha ANSELMI et Bruno GAUJAL.Bâtiment IMAG (amphitheater)Saint-Martin-d'Hères, 38400France -
Séminaire Eyan Castiel
Séminaire Eyan Castiel
–
December 20, 2023Title : Induced idleness leads to deterministic limits in heavy traffic; distributed algorithm and time scale separation
Abstract: In this talk, we’ll analyze a queue-based random-access algorithm where activation and deactivation rates are adapted as functions of local queue lengths. We establish its heavy traffic behavior on a complete interference graph, which turns out to be nonstandard in two respects: (1) the scaling depends on some parameter of the algorithm and is not the N/N^2 scaling usually found in functional central limit theorems; (2) the heavy traffic limit is deterministic. This nonstandard behavior arises from the idleness induced by the distributed nature of the algorithm. We’ll see how time scale separation allows for emergent
Batiment IMAG (Room 306) -
Maxime Gonthier: Scheduling Under Memory Constraint in Task-based Runtime Systems
Maxime Gonthier: Scheduling Under Memory Constraint in Task-based Runtime Systems
–
December 21, 2023Hardware accelerators, such as GPUs, now provide a large part of the computational power used for scientific simulations. GPUs come with their own limited memory and are connected to the main memory of the machine via a bus with limited bandwidth. Scientific simulations often operate on very large data, to the point of not fitting in the limited GPU memory. In this case, one has to turn to out-of-core computing: data are kept in the CPU memory, and moved back and forth to the GPU memory when needed for the computation. This out-of-core situation also happens when processing on multi-core CPUs with limited memory huge datasets stored on disk. In both cases, data movement quickly becomes a performance bottleneck.
Task-based runtime schedulers have emerged as an efficient way to manage large applications. They are in charge of choosing which tasks to assign on which processing unit and in which order they should be processed. During this talk, we aim to present the problem of scheduling for a task-based runtime to improve data locality in an out-of-core setting, in order to reduce data movements.
We designed strategies for both task scheduling and data eviction from limited memories. We implemented them in the StarPU runtime and compared them to existing scheduling techniques in runtime systems. Our strategies achieves significantly better performance when scheduling tasks on multiple GPUs with limited memory, as well as on multiple CPU cores with limited main memory.
We also present work on batch scheduling of IO intensive workloads. Similarly, we used data locality techniques to reduce the average latency of a job.
In a more general approach, we would like to discuss the implications of introducing locality-aware techniques into different domains and the importance of performance visualizations.
Bâtiment IMAG (406) -
[Talk] Felipe Garrido-Lucero
[Talk] Felipe Garrido-Lucero
–
January 10, 2024Who: Felipe Garrido-Lucero
When: Wednesday 10/01, 14:00-15:00
Where: Room 106 and zoom (https://univ-grenoble-alpes-fr.zoom.us/my/mertikop)
What: Fairness Challenges in AI: Insights into Data-set Valuation and Matching Markets
More: Fairness is an increasingly research subject within various computer science sub-disciplines such as machine learning and market design. In this presentation we will discuss two possible fairness challenges: the data-set valuation problem and the price of fairness in bipartite matching markets.
The data-set valuation problem addresses measuring the contributions of agents when collaborating on machine learning tasks. By employing tools from both machine learning and game theory, we model this as a cooperative game and present an approach to approximate the Shapley value of players. The method demonstrates superior performance compared to Monte-Carlo state-of-the-art techniques, supported by theoretical guarantees.
The price of fairness (PoF) quantifies the optimality loss when applying fairness constraints to a problem. Examining the egalitarian PoF in bipartite matching markets, where agents belong to distinct groups, we exploit matroid and geometric tools to characterize fair matchings, optimal matchings, and their intersection. An adversarial analysis reveals that a PoF of 1 is always achievable for two groups, while for a greater number of groups, the PoF can be arbitrarily large.
Bâtiment IMAG (106) -
[Talk] Henri Lefebvre
[Talk] Henri Lefebvre
–
January 19, 2024Who: Henri Lefebvre
When: Friday 19/01, 10:30-11:30
Where: Room 406
What: Exact Methods in Adjustable Robust Optimization
More: This talk will be dedicated to solving complex decision-making problems in which (part of) the input data is not known at decision time. More precisely, we will discuss problems in which a two-stage decision phase is at stake. In the first stage, decisions must be made without complete information about the input parameters, while in the second stage, additional decisions can be informed by revealed information regarding the parameters. More specifically, we will dive into the theoretical derivation and practical use of exact methods for such problems with particular emphasis on the challenging setting in which the second-stage decisions are solution of a mixed-integer non-linear problem.
Bâtiment IMAG (442)
