Experimental Design in Dynamical System Identification: A Bandit-Based Active Learning Approach

Abstract : This study focuses on dynamical system identification, with the reverse modeling of a gene regulatory network as motivating appli-cation. An active learning approach is used to iteratively select the most informative experiments needed to improve the parameters and hidden variables estimates in a dynamical model given a budget for experiments. The design of experiments under these budgeted resources is formalized in terms of sequential optimization. A local optimization criterion (re-ward) is designed to assess each experiment in the sequence, and the global optimization of the sequence is tackled in a game-inspired setting, within the Upper Confidence Tree framework combining Monte-Carlo tree-search and multi-armed bandits. The approach, called EDEN for Experimental Design for parameter Estimation in a Network, shows very good performances on several re-alistic simulated problems of gene regulatory network reverse-modeling, inspired from the international challenge DREAM7.
Document type :
Conference papers
Complete list of metadatas

Cited literature [18 references]  Display  Hide  Download

https://hal.inria.fr/hal-01109775
Contributor : Marc Schoenauer <>
Submitted on : Tuesday, January 27, 2015 - 6:44:54 PM
Last modification on : Monday, December 9, 2019 - 5:24:06 PM
Long-term archiving on: Saturday, September 12, 2015 - 6:35:19 AM

File

dalcheSebagECML2014-author.pdf
Files produced by the author(s)

Identifiers

Citation

Artémis Llamosi, Adel Mezine, Florence d'Alché-Buc, Véronique Letort, Michèle Sebag. Experimental Design in Dynamical System Identification: A Bandit-Based Active Learning Approach. Machine Learning and Knowledge Discovery in Databases - Part II, Sep 2014, Nancy, France. pp.306 - 321, ⟨10.1007/978-3-662-44851-9_20⟩. ⟨hal-01109775⟩

Share

Metrics

Record views

1402

Files downloads

644