June, Friday 22th

8h30 - 17h (room 1014, 660 building) (see location)

Isabelle Guyon's group seminar day: MEDI-CHAL / L2RPN

MORNING: MEDI-CHAL

8h30: Isabelle Guyon: Introduction
9h : Saloni Dash, "A study of Privacy Metrics for synthetic data"
9h30 : Diviyan Kalainathan, "SAM: Structural Agnostic Model"
10h : Break
10h30 : Thomas Gerspacher, "Using synthetic medical data to create challenges"
11h : Adrien Pavao, "Multi-field categorical data encodings"
11h30 : Ritik Dutta. “Metrics of comparison of distributions”.

12h00-14h: Lunch

AFTERNOON: L2RPN

14h00 : Isabelle Guyon: Introduction
14h30: Marvin Lerousseau, "Design and implementation of an environment for Learning to Run a Power Network (L2RPN)”
15h00: Kimang Khun: Preliminary RL results on L2RPN
15h30: Break
16h00-17h00: Discussion.

17h00: Adjourn

==

Medi-chal: This is a project in collaboration with RPI, New-York, USA. The goal of the project is to teach data science to medical students. To that end we are creating mini-challenges, starting from one on classification and feature selection and one on survival analysis. To protect the confidentiality of real medical data, we are generating artificial (but realistic) data using trained data generating models. This poses the problem of the tradeoff between protecting privacy and obtained realistic data. Several methods have been tried, including baseline methods such as Parzen windows, multiple-imputations of (purposely introduced) missing values, copulas, and GANs. We also tried the in-house SAM causal model.

L2RPN (Learning to Run a Power Network): This project is in collaboration with RTE. The goal of the project is to organize a reinforcement learning challenge to solve the problem of controlling the French high voltage electricity transmission grid. More specifically, all equipments must be maintained in security at all time while supply and demand of electricity are balanced. When a subset of lines are out of service (for maintenance and/or because of incidents cause by storms or other weather conditions), some lines remaining in service may be overloaded. The role of dispatchers (operators) is to modify the topology of the grid to route electricity in a way that no line overheats (thus risking to also break and eventually causing a cascaded failure leading t a blackout). We are creating a simplified version of this problem in form of a game that lends itself to being solved with RL.




Contact: guillaume.charpiat at inria.fr
All TAU seminars: here