Solving adaptive sampling problems in graphical models using Markov decision process

Archive ouverte

Bonneau, Mathieu | Dubois Peyrard, Nathalie | Sabbadin, Régis

Edité par CCSD -

In environmental management problems, decision should ideally rely on knowledge of the whole system. However, due to limited budget, in practice only a small part of the system is sampled and the complete system state is reconstructed from the sampled observations. In this article we consider the situation where the biological system under study is structured and can be modeled as a graphical model. Optimal sampling in such models still raises some methodological questions, like adaptive sampling, or the measure of the quality of a sample in terms of quality of reconstruction. Here, we present a way to formalise these two questions. The sample is chosen as the one which maximises the expected utility of information brought by the observations minus the sample cost.The utily is derived from the notion of Maximum a Posteriori. This problem is known to be NP-hard. We present how to modelit as a Markov decision process in order to build approximate solution methods based on Reinforcement Learning.

• Description
• Sujet/Public
• Outil
• Outil d'anticipation
• Mémoire et thèse
• Gestion