2013-09-19: Joseph Dureau & Thierry Dumont (room 01 of ENSCP, in front of IHP)

Post date: 26-Aug-2013 09:04:00

The Public Library of Models: an open source project towards social modeling

Taking the example of epidemiology, the vagaries of mathematical formulations and practical implementations create frictions that prevent modelers from sharing their work, comparing their results, and making them easily actionable to serve decision-making. The objective of the plom-pipe library is to reduce this technical friction, providing a high-level grammar for the definition of a wide class of state-space models. For models falling within this grammar, corresponding simulation code and Bayesian inference algorithms can be automatically generated in parallelisable C. The definition of standard mathematical formulations and implementations allows wide and systematic comparison between models, facilitating model selection. In addition, inference pipelines can be constructed by leveraging approximate and exact mathematical formulations of the model, for an efficient and automated exploration of posterior probability densities. Lastly, this standardisation process has permitted the development of a web portal for modelers to publish, share, and review their work in real time.

* 4.15pm: Thierry Dumont

Indoor localization : a simultaneous localization and mapping algorithm

Performing indoor localization is a challenging task due to the unpredictable behaviour of radio waves in confined environments. The accuracy strongly relies on the quality of the propagation maps estimation. Usually, such estimation is performed deterministically or using an off-line measurement campaign. However, these techniques do not take into account the environmental dynamics responsible for accuracy degradation. I will present an on-line estimation procedure developed by S. Le Corff and

myself. The procedure relies on a semi-parametric propagation model that considers the influence of the environment on the waves propagation. The inference task is performed on-line using the information sequentially thanks to an on-line Expectation-Maximisation algorithm. The performance of the algorithm is illustrated using both simulated data and a true data set.