Sites Inria

Version française

SEQUEL Research team

Sequential Learning

Team presentation

Lots of artificial systems (either software agent, or hardware robot) obtain their data sequentially, along time. For instance, these data may be web pages that are created, or modified, or removed; these data may be sensor measures gathered along time, either passively, or actively by a system having a feedback on its enviromnent. From these data, these systems extract information that may be used to detect objects (classification problem), or to estimate the parameters of a process (estimation problem), or to interact with its environment (sequential decision problem). Sometimes, the amount of data is also so huge that the only way to process it is to split the data and process each chunk at a time. In all cases, we expect that at each instant, the system is able to provide an adequate response. Typically, the environment is stochastic, and may be non stationary.

The goal of the project-team is to develop concepts and algorithms that are able to process efficiently, and with a known and controlled accuracy, these tasks of sequential learning.

The applications are potentially very numerous. Currently, we work on trecommendation systems. We have had a very significant contribution to go, and more generally games, with the award winning Crazy Stone software.

Research themes

  • sequential learning
  • decision making under uncertainty
  • bandit problems; exploration/exploitation dilemma
  • reinforcement learning

International and industrial relations

  • industrielle : Google, Deepmind, Facebook AI Research, Critéo, Renault, Le Livre Scolaire, ...
  • scientifique : U. Mc Gill (Canada), U. Leoben (Austria), LIP 6, MILA (montréal), ...

Keywords: Machine learning Reinforcement learning Decision making under uncertainty Bandits Monte Carlo methods Statistical learningtheory Kernel methods