An Adaptive Sampling Strategy to Accelerate Markov Chains Monte Carlo
H. Hamdi, I. Couckuyt, M. Costa Sousa, T. Dhaene and C.R. Clarkson
Event name: 79th EAGE Conference and Exhibition 2017
Session: Uncertainty and History Matching - From Exploration to Production
Publication date: 12 June 2017
Info: Extended abstract, PDF ( 1.13Mb )
Price: € 20
Markov chain Monte Carlo (McMC) is a technique to sample the posterior distributions in order to quantify the uncertainty and update our knowledge about the model parameters. However, this technique is computationally costly. Therefore, it is frequently combined with a cheap-to-evaluate proxy model. Building an efficient and accurate proxy model using one-shot sampling method (such as Latin-Hypercube) is a challenging task as we do not know a prior the best sample locations and the number of required samples to tune our proxy model. In this paper, we present a novel adaptive sampling techniques using LOLA-Voronoi and Expected Improvement techniques to sequentially update our proxy (here ordinary Kriging model) using best sampling locations and density. This algorithm is used to balance between the exploration and exploitation to create an accurate proxy model for the IC fault model’s misfit function. The results show a large improvement comparing to the one-shot Latin-Hypercube design. The built proxy is then employed to reduce the computational cost of McMC process to find the posterior distributions of the model parameters.