On internal consistency, conditioning and models of uncertainty
|
|
- Christiana Juliana Williams
- 6 years ago
- Views:
Transcription
1 On internal consistency, conditioning and models of uncertainty Jef Caers, Stanford University Abstract Recent research has been tending towards building models of uncertainty of the Earth, not just building a single (or few) detailed Earth models. However, just as any model, models of uncertainty often need to be constrained/conditioned to data to have any prediction power or be useful in decision making. In this presentation, I propose the concept of internal consistency as a scientific basis to study prior (unconditional) and posterior (conditional) models of uncertainty as well as the various sampling techniques involved. In statistical science, internal consistency is the extent to which tests or procedures assess the same characteristic, skill or quality. In the context of uncertainty, I will therefore define internal consistency as the degree to which sampling methods honor the relationship between the unconditional model of uncertainty (prior) and conditional model of uncertainty (posterior) as specified under a (subjectively) chosen theory (for example: Bayes rule). The tests performed are then various different ways of sampling from the same (conditional or unconditional) distributions. If these distributions are related to each other via a theory, then such tests should yield similar results. I propose various such tests using Bayes rule as the theory. A first test is simply to generate unconditional models, extract data from them using a forward model and generate conditional models from this randomized data. Internal consistency with Bayes rule would mean that both sets of conditional and unconditional models span the exact same space of uncertainty, simply because the data spans the uncertainty in the prior. I show that this not true for a number of popular conditional stochastic modeling methods: sequential simulation with hard data, gradual deformation and ensemble Kalman filters for solving inverse problems. I also show that in some cases lack of internal consistency leads to a considerable artificial reduction of (conditional) uncertainty that may have important consequences if such models are used for prediction purposes. A case involving predicting flow behavior is presented. Finally, I offer some discussion on the importance of internal consistency in practical applications and introduce some novel approaches to conditioning that are internally consistent as well as computationally efficient.
2 Introduction There has been a shift in recent year in building models of uncertainty instead of just building models. What do I mean by that? In 3D modeling, one is interested in creating a 3D gridded model of the Earth representing the data and geological understanding of the spatial structures and components. In a way such a model is a model of what we know. In modeling uncertainty we are interested in covering the uncertainty about a certain response evaluated on these models such as flow simulations, we are not merely interested in just building a 3D model. Such as model is a model of what we don t know. Therefore modeling uncertainty is more than just cranking the random numbers of the same model a few more times (Caers, 2011), it requires a conceptual change in thinking. What is our state of knowledge, what is our lack of understanding, how do we quantify this? Often critical parameters in this uncertainty model need to be identified because too many parameters are uncertain for any model of uncertainty to be useful. Just like 3D model that are constrained to data, models of uncertainty need to be constrained to data. But what does this mean? Does it simply mean that every model in the set of 3D models generated needs to match the data in the same fashion? To judge such conditioning, we introduce the concept of internal consistency. In order not to invent yet another term, I borrowed this notion from statistical science where this refers to the extent to which tests or procedures assess the same characteristic, skill or quality. It is a measure of the precision between the observers or of the measuring instruments used in a study. For example, a researcher designs a questionnaire to find out about college students' dissatisfaction with a particular textbook. Analyzing the internal consistency of the survey items dealing with dissatisfaction will reveal the extent to which items on the questionnaire focus on the notion of dissatisfaction. In this paper we will not necessarily follow this literal interpretation, but will focus on the fact that if two procedures test the same skill or property, then their scores should be similar. The property being studied here is the conditioning of various methods in reservoir modeling, be it to well-log or production data. At first, we will focus on a simple test, namely that if models of uncertainty are conditioning to a random set of data, then this conditional model of uncertainty should be the same the unconditional model of uncertainty. If that is not the case than the conditioning method is internally consistent with the theory that links conditional and unconditional models. We will see for example conditioning techniques that perfectly match the data are in fact internally inconsistent. Internal consistency for Earth models Two schools of thought Internal consistency requires a theory, hypothesis or objective, therefore we call it internal. There is no such thing as absolute internal consistency or external consistency. In conditioning we can use
3 Bayes rule as such a theorem. But it should be stated that we do not need to use Bayes rule, we could invent other rules, as long as we stay consistent with those rules, then, we know what we are doing and are following a scientifically rigorous path. Bayes rule is very simple, f MD ( m d) f D M ( d m) f ( m) f () d D M where M is the random vector describing a gridded Earth model, D is the random vector describing the data outcomes. Bayes rule basically states the relationship between the prior or unconditional model of uncertainty f M and the conditional or posterior model of uncertainty f M D. Once you choose the prior and the likelihood function, the posterior is fixed; you cannot choose it any longer independently without being internally inconsistent with Bayes rule. P(m) P(m d) d RN Sampler Explicit theory m (1) m (2) m (3)... Bayesian view Model of spatial continuity statistics d RN algorithm Implicit theory m (1) m (2) m (3)... frequentist view Figure 1: flowchart depiction of two schools of thought Generally two schools of thought have prevailed in geostatistics, see Figure 1. The Bayesian school of thinking prescribed that one should explicitly state the prior and likelihood, then use Bayes rule to determine the posterior and then sample as accurately as possible from this posterior distribution. By accurate sampling, we mean uniform sampling. This way of thinking is very rigorous, but also very cumbersome. First, very few explicit multi-variate distributions are known. Secondly, parameterizing and
4 stating explicitly parameters for these distributions is tedious and as a result simple models, such as multi-gaussian with simplified parameters assumptions (such as homoscedasticity in the variance) are assumed. Last but not least, sampling methods such as McMC are impractical if the data-model relationship is complex or CPU demanding to evaluate. But the Bayesian view is internally consistent. In juxtaposition, there is a more frequentist approach to modeling, namely, that any set of 3D Earth models represents a model of uncertainty, whether or not these models are conditional or unconditional. In their view, there is no need for specifying any distribution functions explicitly. As long as one can create models, one is fine. Any algorithm can be viewed as a model and this algorithm can be conditional or unconditional. Bayes rule is used in various ways, for example to constrain to data, but it is generally not used in an explicit way to link conditional and unconditional models of uncertainty. But are models created in this way internally consistent? Let s consider an example. Testing for internal consistency We design a simple test for internal consistency between conditioning mechanism, theory and models of uncertainty. Recall that as theory we chose for Bayes rule, a subjective choice, but a choice nonetheless. Our test works as follows and works for both Bayesian and Frequentist views. We assume that the data-model relationship is given by a forward model, namely d g( m ) In the Bayesian world the test would be 1. Sample m from f M 2. Generate d from d=g(m) 3. Specify the likelihood f D M and therefore the posterior 4. Sample now from f M D (m d) For the frequentist view, one has the same thing, but different ways of expressing it in practice: 1. Generate an unconditional simulation m using an unconditional algorithm 2. Generate d from d=g(m) 3. Generate a conditional simulation m d using the conditional algorithm What is the purpose here? In repeating this workflow, we obtain multiple conditional Earth models. The distribution of these conditional Earth model should be exactly the same as the unconditional or prior model, indeed, f ( m d) f ( d) dd f ( d m) f ( m) dd f ( m) f ( d m) dd f ( m) M D D D M M M D M M d d d (1)
5 This makes sense, since we randomize the data in such a way that it is consistent with the prior. Note that in this derivation we used our theory, namely Bayes rule. i= Figure 2: outline of the simple Boolean model A simple example To illustrate this internal consistency test, we use a simple but perhaps baffling example of what can go wrong. Consider a 1D grid, see Figure 1. The prior model is a simple Boolean model. The Boolean model consist placing exactly five objects on this line, where each object is drawn from a given uniform distribution with length [minl, maxl]. The objects can overlap. As data, we consider the exact observation of absence/presence of an object at the middle location, see Figure 1. We describe the random variable a A(i), where A(i)=1 indicates that the object is present, A(i)=0 that the object is absent. Consider a simple conditioning method, i.e. a method for generating conditional realizations. Clearly we have two case, either A(50)=1 or A(50)=0. If the first case would occur, we generate a conditional model as follows: 1. Draw an object with certain length from the uniform distribution 2. Put it uniformly around the conditioning location Generate the four remaining objects When A(50), we do the following 1. Draw the length of a single object 2. Put it uniformly at those locations that will not violate the conditioning data A(50)=0 3. Repeat this till you have 5 objects Clearly this will generate conditional simulations, and it appears, follow the Boolean model. Not quite. Let s run our consistency test. First we need to know the marginal P(A(i)=1). This is easy, we simple generate 1000 unconditional models and calculate their ensemble average. To run our test, we simply generate N conditional models with A(50)=1, N conditional models with A(50)=0, calculate their ensemble averages and average them according to the marginal, namely
6 EA EA P( A 1) EA (1 P( A 1)) 1 0 If the conditioning method were to be internally consistent, one should obtain that EA PA ( 1) that is, get back the marginal as stated in Eq (1). Figure 3 shows that this is not the case, the marginal is shown in red, while the result from our test is shown in blue. The conclusion is that this method is not internally consistent. What happened? To further analyze this puzzling result, we use a sampler in the rigorous Bayesian sense that is known to be exact, the rejection sampler. In rejection sampling we simply generate a model and accept it is it matches the data, reject it else. Consider first rejection sampling when A(50)=0. In Figure 4 we plot the ensemble average (conditional mean) of 1000 rejection sampler results together the ensemble average of the simple conditioning method. We get a perfect match. The result is however different for the case A(50)=1,see Figure 4. Indeed it seems that the average size of objects generated with our technique is too small. This makes sense because of the fact that an observation of on object is more likely when the object is large than when it is small, a fact that was not considered in our simple conditioning method P( m) P(A=1) D grid d P( m d)p( d)dd Figure 3: results of the internal consistency test
7 P(A=0 A(50)=1) P(A=1 A(50)=1) Rejection sampler Rejection sampler Naïve conditioning Naïve conditioning 1D grid 1D grid Figure 4: comparing rejection sampler with simple conditioning for both cases A(50)=0 and A(50)=1 Training image Grid with single hard data Channel sand Background mud Figure 5: example TI and single hard data More examples In the current and previous SCRF reports we have provided several example illustrations of this internal consistency property. I will briefly summarize the results.
8 Rejection sampler unconditional sequential simulation Conditional sequential simulation Random path Figure 6: ensemble ab=verage of the rejection sampler and conditional MPS using snesim Conditional MPS simulation The conditional MPS simulation algorithms often differ from the unconditional ones, just as in the Boolean example, this can lead to internal inconsistency. Consider the simple example shown in Figure 5. A single hard conditioning data indicating sand is located in the center. A training image is given on the left of simple sinuous channels. Consider now conditioning first using rejection sampler. 150 models are created that match the data. The rejection sampler uses the unconditional version of snesim. The ensemble average is provided on the left in Figure 6. Next 150 models are created using the same snesim algorithm, but now the conditional version. Clearly the ensemble average in Figure 6 differs from the rejection sampler, meaning that there exists an internal consistency problem between conditional and unconditional snesim. Where does this problem occur? Since snesim works on multi-grids, the single hard data needs to be relocated to the nearest coarse grid node. This data re-location does not occur when performing unconditional simulation; hence this is the source of the discrepancy. In the work of Honarkhah (2011), this problem is resolved using a different data relocation algorithm, see his work for details. Figure 7 show this his dispat code indeed does have results comparable to rejection sampling, even in cases with complex data configurations as opposed to the simple data relocation implemented in snesim.
9 Figure 7: results from dispat History matching by ensemble Kalman filters Ensemble Kalman filters (EnKf) have recently been popular in researching methods for obtaining multiple history matched models. In this method a set of initial or prior reservoir models is generated. Next, a first time step of the flow simulation is executed and dynamic variables are calculated. This initial set of models is then updated (linearly) as a whole using the difference between the field data and the response of the initial set, as well as the covariance matrix between the static and dynamic variables. The theory requires that the models a multi-gaussian and the relationship between data and model is linear (or almost linear). This update is repeated until the last time step of the flow simulation. Consider an example in Figure 8. A simple injector and producer configuration is shown on a 31x31 grid. As prior
10 model, we have a training image, see Figure 9, from which a set of initial reservoir models can be generated, see Figure 10. To apply the EnKf method to these clearly non-gaussian fields, we use the metric ensemble Kalman filter (Park, 2011), an adaptation of the ensemble Kalman filter performed after kernel transformation. Figure shows that success is obtained by generating 30 models that reproduce the training image patterns as well as match the data. So, to the eye of the innocent bystander, everything seems perfect. Consider however generating 30 history matched models using the rejection sampler. When then comparing the conditional variance of the permeability fields generated using these two techniques, see Figure 13, we notice the low variance of the ensemble Kalman filter as compared to the rejection sampler. Clearly the linear and Gaussian hypothesis of the Kalman filter lead to internal inconsistency. The data Figure 8: setup problem for the EnKf Figure 9: training image
11 In the work of Park (2011), a different approach is taken that does not require linearity, nor Gaussianity. His distance-based approach has been tested on the same example resulting in Figure 13. Clearly, while not yet perfect, he achieves a much greater randomization than the EnKf, while matching the data equally well. Unconditional simulation Figure 10: unconditional simulation Some posterior models obtained by EnKf Figure 11: conditional simulation
12 Rejection sampler forward simulations EnKf 30 forward simulations Figure 12: history matching with rejection sampler and Enkf Rejection sampler EnKf Distance methods Figure 13: comparing conditional variance
13 The data The model Rock type and boundary of surfaces Surface-based model Figure 14: 9 wells of a real case and the surface-based model History matching by means of optimization The shortcoming of the EnKf method is that it is basically a kriging-type approach (linear updating, covariances), hence only works for posterior determination if that posterior is multi-gaussian. Once can therefore question how well other optimization technique work in terms of internal consistency when applied to the history matching problem, since basically EnKf, as kriging, is an optimization technique. How well do we match the 8 wells? How well do we predict the one well? Bias Variance Figure 15: (top) matching performance, (bottom)prediction performance
14 Consider therefore a more complex example, studied in Bertoncello (2011). In his work, a complex surface-based forward model is built, see Figure 14, based on various input parameters such as length and height of the lobes, origin location, migration and progradation statistics as well as several rules related to the deposition and erosion of these bodies. Spatial uncertainty is modeled by placing the lobe surfaces in various different positions. A complex but geologically realistic model can be created. In order to fit such a model to data, such as well data, one can execute an iterative trial-and-error type optimization algorithm that modifies the lobe parameters and placement such that some objective function quantifying the mismatch between data and model is minimized. In order to check how predictive such optimization is, consider a realistic case outlined in Bertoncello (2011). A surface-based model is matched iteratively to 8 wells, with the aim of predicting the outcome of a 9 th well in a real-case dataset. Figure 15 shows that as the iteration proceeds, the match is getting better. Several models were matched, each starting from a different initial solution, providing an envelope of mismatches. All models reached a good match as long as the optimization is ran long enough. How well do these models predict the 9 th well. The prediction is good up to a certain amount of matching. Clearly when the iterations are run a long time, the predictions will start to deteriorate, meaning that the solutions spread an uncertainty space that has become too narrow and also biased. Focusing too much on matching data may therefore create poor models of uncertainty. There may be various reasons for this in this example. The forward model may not accurately capture reality, hence any strict matching leads to a bias. Secondly, optimization methods tend to provide a too narrow space of uncertainty. Does it matter? One can wonder whether this sudden focus on internal consistency really matters. Consider again the MPS conditional simulation and consider now predicting flow in a neighboring producer well, see Figure 16. Clearly the uncertainty in terms of flow can be highly affected. Nevertheless, one should have a broader discussion on internal consistency than this simple example. If we return to Bayes rule than the role of the prior becomes important. Often we have a good handle on the likelihood, that is, how well we should match the data. The problem in modeling uncertainty often lies in the prior. Clearly in the above examples, the data was matched, but it was matched incorrectly. In previous years, we put a lot of emphasis on geological consistency. This geological consistency issue is not the cause of the incorrect matching. All models reproduce the prior statistics and the data. But the posterior has become inconsistent with the prior. Does this matter? It matters if considerable effort has been put in constructing the prior, such as for example the case with multiple training images. If however the prior is multi-gaussian, then, to my opinion, there is no need to be consistent with it since it is already a fabricated model that is not very in tune with reality.
15 % water produced Grid with single hard data injector Rejection sampler Conditional simulation P producer P Time Figure 16: the consequence in terms of flow prediction of internal inconsistency References Bertoncello, A., Conditioning of Surface-Based Models to Wells and Thickness Maps. PhD dissertation, Stanford University. Caers, J., Modeling Uncertainty in the Earth Sciences. Wiley-Blackwell. 250p. Honarkhah, M., Stochastic Simulation of Patterns Using Distance-based Pattern Modeling. PhD dissertation, Stanford University. Park, K, Modeling Uncertainty in Metric Space, PhD dissertation, Stanford University.
A workflow to account for uncertainty in well-log data in 3D geostatistical reservoir modeling
A workflow to account for uncertainty in well-log data in 3D geostatistical reservoir Jose Akamine and Jef Caers May, 2007 Stanford Center for Reservoir Forecasting Abstract Traditionally well log data
More informationAdaptive spatial resampling as a Markov chain Monte Carlo method for uncertainty quantification in seismic reservoir characterization
1 Adaptive spatial resampling as a Markov chain Monte Carlo method for uncertainty quantification in seismic reservoir characterization Cheolkyun Jeong, Tapan Mukerji, and Gregoire Mariethoz Department
More informationConditioning a hybrid geostatistical model to wells and seismic data
Conditioning a hybrid geostatistical model to wells and seismic data Antoine Bertoncello, Gregoire Mariethoz, Tao Sun and Jef Caers ABSTRACT Hybrid geostatistical models imitate a sequence of depositional
More informationFMA901F: Machine Learning Lecture 3: Linear Models for Regression. Cristian Sminchisescu
FMA901F: Machine Learning Lecture 3: Linear Models for Regression Cristian Sminchisescu Machine Learning: Frequentist vs. Bayesian In the frequentist setting, we seek a fixed parameter (vector), with value(s)
More informationSampling informative/complex a priori probability distributions using Gibbs sampling assisted by sequential simulation
Sampling informative/complex a priori probability distributions using Gibbs sampling assisted by sequential simulation Thomas Mejer Hansen, Klaus Mosegaard, and Knud Skou Cordua 1 1 Center for Energy Resources
More informationModeling Uncertainty in the Earth Sciences Jef Caers Stanford University
Modeling spatial continuity Modeling Uncertainty in the Earth Sciences Jef Caers Stanford University Motivation uncertain uncertain certain or uncertain uncertain Spatial Input parameters Spatial Stochastic
More informationIterative spatial resampling applied to seismic inverse modeling for lithofacies prediction
Iterative spatial resampling applied to seismic inverse modeling for lithofacies prediction Cheolkyun Jeong, Tapan Mukerji, and Gregoire Mariethoz Department of Energy Resources Engineering Stanford University
More informationA009 HISTORY MATCHING WITH THE PROBABILITY PERTURBATION METHOD APPLICATION TO A NORTH SEA RESERVOIR
1 A009 HISTORY MATCHING WITH THE PROBABILITY PERTURBATION METHOD APPLICATION TO A NORTH SEA RESERVOIR B. Todd HOFFMAN and Jef CAERS Stanford University, Petroleum Engineering, Stanford CA 94305-2220 USA
More informationAntoine Bertoncello, Hongmei Li and Jef Caers
Antoine Bertoncello, Hongmei Li and Jef Caers model created by surface based model (Bertoncello and Caers, 2010) A conditioning methodology have been developed and tested on a model with few parameters
More informationMultiple Point Statistics with Multiple Training Images
Multiple Point Statistics with Multiple Training Images Daniel A. Silva and Clayton V. Deutsch Abstract Characterization of complex geological features and patterns has been one of the main tasks of geostatistics.
More informationB. Todd Hoffman and Jef Caers Stanford University, California, USA
Sequential Simulation under local non-linear constraints: Application to history matching B. Todd Hoffman and Jef Caers Stanford University, California, USA Introduction Sequential simulation has emerged
More informationA Parallel, Multiscale Approach to Reservoir Modeling. Omer Inanc Tureyen and Jef Caers Department of Petroleum Engineering Stanford University
A Parallel, Multiscale Approach to Reservoir Modeling Omer Inanc Tureyen and Jef Caers Department of Petroleum Engineering Stanford University 1 Abstract With the advance of CPU power, numerical reservoir
More informationJoint quantification of uncertainty on spatial and non-spatial reservoir parameters
Joint quantification of uncertainty on spatial and non-spatial reservoir parameters Comparison between the Method and Distance Kernel Method Céline Scheidt and Jef Caers Stanford Center for Reservoir Forecasting,
More informationGeostatistical Reservoir Characterization of McMurray Formation by 2-D Modeling
Geostatistical Reservoir Characterization of McMurray Formation by 2-D Modeling Weishan Ren, Oy Leuangthong and Clayton V. Deutsch Department of Civil & Environmental Engineering, University of Alberta
More informationRotation and affinity invariance in multiple-point geostatistics
Rotation and ainity invariance in multiple-point geostatistics Tuanfeng Zhang December, 2001 Abstract Multiple-point stochastic simulation of facies spatial distribution requires a training image depicting
More informationSupport vector machines
Support vector machines Cavan Reilly October 24, 2018 Table of contents K-nearest neighbor classification Support vector machines K-nearest neighbor classification Suppose we have a collection of measurements
More informationGeostatistical modelling of offshore diamond deposits
Geostatistical modelling of offshore diamond deposits JEF CAERS AND LUC ROMBOUTS STANFORD UNIVERSITY, Department of Petroleum Engineering, Stanford, CA 94305-2220, USA jef@pangea.stanford.edu TERRACONSULT,
More informationRecap: Gaussian (or Normal) Distribution. Recap: Minimizing the Expected Loss. Topics of This Lecture. Recap: Maximum Likelihood Approach
Truth Course Outline Machine Learning Lecture 3 Fundamentals (2 weeks) Bayes Decision Theory Probability Density Estimation Probability Density Estimation II 2.04.205 Discriminative Approaches (5 weeks)
More informationExploring Direct Sampling and Iterative Spatial Resampling in History Matching
Exploring Direct Sampling and Iterative Spatial Resampling in History Matching Matz Haugen, Grergoire Mariethoz and Tapan Mukerji Department of Energy Resources Engineering Stanford University Abstract
More informationCS 229 Midterm Review
CS 229 Midterm Review Course Staff Fall 2018 11/2/2018 Outline Today: SVMs Kernels Tree Ensembles EM Algorithm / Mixture Models [ Focus on building intuition, less so on solving specific problems. Ask
More informationMissing Data. Where did it go?
Missing Data Where did it go? 1 Learning Objectives High-level discussion of some techniques Identify type of missingness Single vs Multiple Imputation My favourite technique 2 Problem Uh data are missing
More informationB002 DeliveryMassager - Propagating Seismic Inversion Information into Reservoir Flow Models
B2 DeliveryMassager - Propagating Seismic Inversion Information into Reservoir Flow Models J. Gunning* (CSIRO Petroleum) & M.E. Glinsky (BHP Billiton Petroleum) SUMMARY We present a new open-source program
More informationMarkov Bayes Simulation for Structural Uncertainty Estimation
P - 200 Markov Bayes Simulation for Structural Uncertainty Estimation Samik Sil*, Sanjay Srinivasan and Mrinal K Sen. University of Texas at Austin, samiksil@gmail.com Summary Reservoir models are built
More informationCONDITIONAL SIMULATION OF TRUNCATED RANDOM FIELDS USING GRADIENT METHODS
CONDITIONAL SIMULATION OF TRUNCATED RANDOM FIELDS USING GRADIENT METHODS Introduction Ning Liu and Dean S. Oliver University of Oklahoma, Norman, Oklahoma, USA; ning@ou.edu The problem of estimating the
More informationRegularization and model selection
CS229 Lecture notes Andrew Ng Part VI Regularization and model selection Suppose we are trying select among several different models for a learning problem. For instance, we might be using a polynomial
More informationMulticollinearity and Validation CIVL 7012/8012
Multicollinearity and Validation CIVL 7012/8012 2 In Today s Class Recap Multicollinearity Model Validation MULTICOLLINEARITY 1. Perfect Multicollinearity 2. Consequences of Perfect Multicollinearity 3.
More informationPrograms for MDE Modeling and Conditional Distribution Calculation
Programs for MDE Modeling and Conditional Distribution Calculation Sahyun Hong and Clayton V. Deutsch Improved numerical reservoir models are constructed when all available diverse data sources are accounted
More informationOutline. Data Association Scenarios. Data Association Scenarios. Data Association Scenarios
Outline Data Association Scenarios Track Filtering and Gating Global Nearest Neighbor (GNN) Review: Linear Assignment Problem Murthy s k-best Assignments Algorithm Probabilistic Data Association (PDAF)
More informationAn Introduction to Markov Chain Monte Carlo
An Introduction to Markov Chain Monte Carlo Markov Chain Monte Carlo (MCMC) refers to a suite of processes for simulating a posterior distribution based on a random (ie. monte carlo) process. In other
More informationSurface-based model conditioning using an hybrid optimization: methodology and application
Surface-based model conditioning using an hybrid optimization: methodology and application Antoine Bertoncello, Jef Caers, Hongmei Li and Tao Sun Department of Energy Resources Engineering Stanford University
More informationGeostatistics on Stratigraphic Grid
Geostatistics on Stratigraphic Grid Antoine Bertoncello 1, Jef Caers 1, Pierre Biver 2 and Guillaume Caumon 3. 1 ERE department / Stanford University, Stanford CA USA; 2 Total CSTJF, Pau France; 3 CRPG-CNRS
More informationSupplementary Figure 1. Decoding results broken down for different ROIs
Supplementary Figure 1 Decoding results broken down for different ROIs Decoding results for areas V1, V2, V3, and V1 V3 combined. (a) Decoded and presented orientations are strongly correlated in areas
More informationA Geostatistical and Flow Simulation Study on a Real Training Image
A Geostatistical and Flow Simulation Study on a Real Training Image Weishan Ren (wren@ualberta.ca) Department of Civil & Environmental Engineering, University of Alberta Abstract A 12 cm by 18 cm slab
More informationImprovements in Continuous Variable Simulation with Multiple Point Statistics
Improvements in Continuous Variable Simulation with Multiple Point Statistics Jeff B. Boisvert A modified version of Mariethoz et al s (2010) algorithm for simulating continuous variables using multiple
More information3 Nonlinear Regression
3 Linear models are often insufficient to capture the real-world phenomena. That is, the relation between the inputs and the outputs we want to be able to predict are not linear. As a consequence, nonlinear
More informationHierarchical modeling of multi-scale flow barriers in channelized reservoirs
Hierarchical modeling of multi-scale flow barriers in channelized reservoirs Hongmei Li and Jef Caers Stanford Center for Reservoir Forecasting Stanford University Abstract Channelized reservoirs often
More informationSimulating Geological Structures Based on Training Images and Pattern Classifications
Simulating Geological Structures Based on Training Images and Pattern Classifications P. Switzer, T. Zhang, A. Journel Department of Geological and Environmental Sciences Stanford University CA, 9435,
More informationSCRF. 22 nd Annual Meeting. April 30-May
SCRF 22 nd Annual Meeting April 30-May 1 2009 1 Research Overview CD annual report with papers Presentations 2 Modeling Uncertainty Distance based modeling of uncertainty - Model selection - Inverse modeling
More informationApplying Supervised Learning
Applying Supervised Learning When to Consider Supervised Learning A supervised learning algorithm takes a known set of input data (the training set) and known responses to the data (output), and trains
More informationMonte Carlo for Spatial Models
Monte Carlo for Spatial Models Murali Haran Department of Statistics Penn State University Penn State Computational Science Lectures April 2007 Spatial Models Lots of scientific questions involve analyzing
More informationHistory matching under training-image based geological model constraints
History matching under training-image based geological model constraints JEF CAERS Stanford University, Department of Petroleum Engineering Stanford, CA 94305-2220 January 2, 2002 Corresponding author
More informationMultiple Constraint Satisfaction by Belief Propagation: An Example Using Sudoku
Multiple Constraint Satisfaction by Belief Propagation: An Example Using Sudoku Todd K. Moon and Jacob H. Gunther Utah State University Abstract The popular Sudoku puzzle bears structural resemblance to
More informationCS6375: Machine Learning Gautam Kunapuli. Mid-Term Review
Gautam Kunapuli Machine Learning Data is identically and independently distributed Goal is to learn a function that maps to Data is generated using an unknown function Learn a hypothesis that minimizes
More information2. On classification and related tasks
2. On classification and related tasks In this part of the course we take a concise bird s-eye view of different central tasks and concepts involved in machine learning and classification particularly.
More informationRandom projection for non-gaussian mixture models
Random projection for non-gaussian mixture models Győző Gidófalvi Department of Computer Science and Engineering University of California, San Diego La Jolla, CA 92037 gyozo@cs.ucsd.edu Abstract Recently,
More information11-Geostatistical Methods for Seismic Inversion. Amílcar Soares CERENA-IST
11-Geostatistical Methods for Seismic Inversion Amílcar Soares CERENA-IST asoares@ist.utl.pt 01 - Introduction Seismic and Log Scale Seismic Data Recap: basic concepts Acoustic Impedance Velocity X Density
More informationNearest Neighbor Predictors
Nearest Neighbor Predictors September 2, 2018 Perhaps the simplest machine learning prediction method, from a conceptual point of view, and perhaps also the most unusual, is the nearest-neighbor method,
More informationNaïve Bayes for text classification
Road Map Basic concepts Decision tree induction Evaluation of classifiers Rule induction Classification using association rules Naïve Bayesian classification Naïve Bayes for text classification Support
More informationMachine Learning Lecture 3
Machine Learning Lecture 3 Probability Density Estimation II 19.10.2017 Bastian Leibe RWTH Aachen http://www.vision.rwth-aachen.de leibe@vision.rwth-aachen.de Announcements Exam dates We re in the process
More information3 Nonlinear Regression
CSC 4 / CSC D / CSC C 3 Sometimes linear models are not sufficient to capture the real-world phenomena, and thus nonlinear models are necessary. In regression, all such models will have the same basic
More informationLab 9. Julia Janicki. Introduction
Lab 9 Julia Janicki Introduction My goal for this project is to map a general land cover in the area of Alexandria in Egypt using supervised classification, specifically the Maximum Likelihood and Support
More informationApplication of MPS Simulation with Multiple Training Image (MultiTI-MPS) to the Red Dog Deposit
Application of MPS Simulation with Multiple Training Image (MultiTI-MPS) to the Red Dog Deposit Daniel A. Silva and Clayton V. Deutsch A Multiple Point Statistics simulation based on the mixing of two
More informationReservoir Modeling Combining Geostatistics with Markov Chain Monte Carlo Inversion
Reservoir Modeling Combining Geostatistics with Markov Chain Monte Carlo Inversion Andrea Zunino, Katrine Lange, Yulia Melnikova, Thomas Mejer Hansen and Klaus Mosegaard 1 Introduction Reservoir modeling
More informationRM03 Integrating Petro-elastic Seismic Inversion and Static Model Building
RM03 Integrating Petro-elastic Seismic Inversion and Static Model Building P. Gelderblom* (Shell Global Solutions International BV) SUMMARY This presentation focuses on various aspects of how the results
More informationOverview. Monte Carlo Methods. Statistics & Bayesian Inference Lecture 3. Situation At End Of Last Week
Statistics & Bayesian Inference Lecture 3 Joe Zuntz Overview Overview & Motivation Metropolis Hastings Monte Carlo Methods Importance sampling Direct sampling Gibbs sampling Monte-Carlo Markov Chains Emcee
More informationProject and Production Management Prof. Arun Kanda Department of Mechanical Engineering Indian Institute of Technology, Delhi
Project and Production Management Prof. Arun Kanda Department of Mechanical Engineering Indian Institute of Technology, Delhi Lecture - 8 Consistency and Redundancy in Project networks In today s lecture
More informationRobotics. Lecture 5: Monte Carlo Localisation. See course website for up to date information.
Robotics Lecture 5: Monte Carlo Localisation See course website http://www.doc.ic.ac.uk/~ajd/robotics/ for up to date information. Andrew Davison Department of Computing Imperial College London Review:
More informationModeling Multiple Rock Types with Distance Functions: Methodology and Software
Modeling Multiple Rock Types with Distance Functions: Methodology and Software Daniel A. Silva and Clayton V. Deutsch The sub division of the deposit into estimation domains that are internally consistent
More informationNearest Neighbor Classification
Nearest Neighbor Classification Professor Ameet Talwalkar Professor Ameet Talwalkar CS260 Machine Learning Algorithms January 11, 2017 1 / 48 Outline 1 Administration 2 First learning algorithm: Nearest
More informationIntegration of Geostatistical Modeling with History Matching: Global and Regional Perturbation
Integration of Geostatistical Modeling with History Matching: Global and Regional Perturbation Oliveira, Gonçalo Soares Soares, Amílcar Oliveira (CERENA/IST) Schiozer, Denis José (UNISIM/UNICAMP) Introduction
More informationPattern Recognition ( , RIT) Exercise 1 Solution
Pattern Recognition (4005-759, 20092 RIT) Exercise 1 Solution Instructor: Prof. Richard Zanibbi The following exercises are to help you review for the upcoming midterm examination on Thursday of Week 5
More informationAlgorithm-driven and Representation-driven Random Function : A New Formalism for Applied Geostatistics
Algorithm-driven and Representation-driven Random Function : A New Formalism for Applied Geostatistics Alexandre Boucher Dept of Geological and Environmental Sciences Stanford University Abstract This
More informationSTA 4273H: Statistical Machine Learning
STA 4273H: Statistical Machine Learning Russ Salakhutdinov Department of Statistics! rsalakhu@utstat.toronto.edu! http://www.utstat.utoronto.ca/~rsalakhu/ Sidney Smith Hall, Room 6002 Lecture 12 Combining
More informationMCMC Diagnostics. Yingbo Li MATH Clemson University. Yingbo Li (Clemson) MCMC Diagnostics MATH / 24
MCMC Diagnostics Yingbo Li Clemson University MATH 9810 Yingbo Li (Clemson) MCMC Diagnostics MATH 9810 1 / 24 Convergence to Posterior Distribution Theory proves that if a Gibbs sampler iterates enough,
More informationSTA 4273H: Sta-s-cal Machine Learning
STA 4273H: Sta-s-cal Machine Learning Russ Salakhutdinov Department of Statistics! rsalakhu@utstat.toronto.edu! h0p://www.cs.toronto.edu/~rsalakhu/ Lecture 3 Parametric Distribu>ons We want model the probability
More informationStatistical Techniques in Robotics (16-831, F10) Lecture #02 (Thursday, August 28) Bayes Filtering
Statistical Techniques in Robotics (16-831, F10) Lecture #02 (Thursday, August 28) Bayes Filtering Lecturer: Drew Bagnell Scribes: Pranay Agrawal, Trevor Decker, and Humphrey Hu 1 1 A Brief Example Let
More informationComputer Experiments: Space Filling Design and Gaussian Process Modeling
Computer Experiments: Space Filling Design and Gaussian Process Modeling Best Practice Authored by: Cory Natoli Sarah Burke, Ph.D. 30 March 2018 The goal of the STAT COE is to assist in developing rigorous,
More informationFast FILTERSIM Simulation with Score-based Distance Function
Fast FILTERSIM Simulation with Score-based Distance Function Jianbing Wu (1), André G. Journel (1) and Tuanfeng Zhang (2) (1) Department of Energy Resources Engineering, Stanford, CA (2) Schlumberger Doll
More informationOvercompressing JPEG images with Evolution Algorithms
Author manuscript, published in "EvoIASP2007, Valencia : Spain (2007)" Overcompressing JPEG images with Evolution Algorithms Jacques Lévy Véhel 1, Franklin Mendivil 2 and Evelyne Lutton 1 1 Inria, Complex
More informationBayesian Estimation for Skew Normal Distributions Using Data Augmentation
The Korean Communications in Statistics Vol. 12 No. 2, 2005 pp. 323-333 Bayesian Estimation for Skew Normal Distributions Using Data Augmentation Hea-Jung Kim 1) Abstract In this paper, we develop a MCMC
More informationTour-Based Mode Choice Modeling: Using An Ensemble of (Un-) Conditional Data-Mining Classifiers
Tour-Based Mode Choice Modeling: Using An Ensemble of (Un-) Conditional Data-Mining Classifiers James P. Biagioni Piotr M. Szczurek Peter C. Nelson, Ph.D. Abolfazl Mohammadian, Ph.D. Agenda Background
More informationVariability in Annual Temperature Profiles
Variability in Annual Temperature Profiles A Multivariate Spatial Analysis of Regional Climate Model Output Tamara Greasby, Stephan Sain Institute for Mathematics Applied to Geosciences, National Center
More informationA Finite Element Analysis Workflow in Biomechanics. Define the problem and identify FEA as the favorable modeling & simulation strategy
Before you start Understand the Premise of FEA Finite element analysis is a computational modeling & simulation strategy; specifically, a numerical procedure for solving field problems, e.g., stress, by
More informationEquation to LaTeX. Abhinav Rastogi, Sevy Harris. I. Introduction. Segmentation.
Equation to LaTeX Abhinav Rastogi, Sevy Harris {arastogi,sharris5}@stanford.edu I. Introduction Copying equations from a pdf file to a LaTeX document can be time consuming because there is no easy way
More information(Refer Slide Time: 0:51)
Introduction to Remote Sensing Dr. Arun K Saraf Department of Earth Sciences Indian Institute of Technology Roorkee Lecture 16 Image Classification Techniques Hello everyone welcome to 16th lecture in
More informationEnsemble Learning. Another approach is to leverage the algorithms we have via ensemble methods
Ensemble Learning Ensemble Learning So far we have seen learning algorithms that take a training set and output a classifier What if we want more accuracy than current algorithms afford? Develop new learning
More informationMachine Learning Methods in Visualisation for Big Data 2018
Machine Learning Methods in Visualisation for Big Data 2018 Daniel Archambault1 Ian Nabney2 Jaakko Peltonen3 1 Swansea University 2 University of Bristol 3 University of Tampere, Aalto University Evaluating
More informationHigh Resolution Geomodeling, Ranking and Flow Simulation at SAGD Pad Scale
High Resolution Geomodeling, Ranking and Flow Simulation at SAGD Pad Scale Chad T. Neufeld, Clayton V. Deutsch, C. Palmgren and T. B. Boyle Increasing computer power and improved reservoir simulation software
More informationPSU Student Research Symposium 2017 Bayesian Optimization for Refining Object Proposals, with an Application to Pedestrian Detection Anthony D.
PSU Student Research Symposium 2017 Bayesian Optimization for Refining Object Proposals, with an Application to Pedestrian Detection Anthony D. Rhodes 5/10/17 What is Machine Learning? Machine learning
More informationDiscovery of the Source of Contaminant Release
Discovery of the Source of Contaminant Release Devina Sanjaya 1 Henry Qin Introduction Computer ability to model contaminant release events and predict the source of release in real time is crucial in
More informationPractice EXAM: SPRING 2012 CS 6375 INSTRUCTOR: VIBHAV GOGATE
Practice EXAM: SPRING 0 CS 6375 INSTRUCTOR: VIBHAV GOGATE The exam is closed book. You are allowed four pages of double sided cheat sheets. Answer the questions in the spaces provided on the question sheets.
More informationUsing Blast Data to infer Training Images for MPS Simulation of Continuous Variables
Paper 34, CCG Annual Report 14, 212 ( 212) Using Blast Data to infer Training Images for MPS Simulation of Continuous Variables Hai T. Nguyen and Jeff B. Boisvert Multiple-point simulation (MPS) methods
More informationD025 Geostatistical Stochastic Elastic Iinversion - An Efficient Method for Integrating Seismic and Well Data Constraints
D025 Geostatistical Stochastic Elastic Iinversion - An Efficient Method for Integrating Seismic and Well Data Constraints P.R. Williamson (Total E&P USA Inc.), A.J. Cherrett* (Total SA) & R. Bornard (CGGVeritas)
More informationNearest neighbor classification DSE 220
Nearest neighbor classification DSE 220 Decision Trees Target variable Label Dependent variable Output space Person ID Age Gender Income Balance Mortgag e payment 123213 32 F 25000 32000 Y 17824 49 M 12000-3000
More informationCLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS
CLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS CHAPTER 4 CLASSIFICATION WITH RADIAL BASIS AND PROBABILISTIC NEURAL NETWORKS 4.1 Introduction Optical character recognition is one of
More informationCONDITIONING FACIES SIMULATIONS WITH CONNECTIVITY DATA
CONDITIONING FACIES SIMULATIONS WITH CONNECTIVITY DATA PHILIPPE RENARD (1) and JEF CAERS (2) (1) Centre for Hydrogeology, University of Neuchâtel, Switzerland (2) Stanford Center for Reservoir Forecasting,
More information6.001 Notes: Section 8.1
6.001 Notes: Section 8.1 Slide 8.1.1 In this lecture we are going to introduce a new data type, specifically to deal with symbols. This may sound a bit odd, but if you step back, you may realize that everything
More informationPerformance Estimation and Regularization. Kasthuri Kannan, PhD. Machine Learning, Spring 2018
Performance Estimation and Regularization Kasthuri Kannan, PhD. Machine Learning, Spring 2018 Bias- Variance Tradeoff Fundamental to machine learning approaches Bias- Variance Tradeoff Error due to Bias:
More informationEstimation of Item Response Models
Estimation of Item Response Models Lecture #5 ICPSR Item Response Theory Workshop Lecture #5: 1of 39 The Big Picture of Estimation ESTIMATOR = Maximum Likelihood; Mplus Any questions? answers Lecture #5:
More informationHIERARCHICAL SIMULATION OF MULTIPLE-FACIES RESERVOIRS USING MULTIPLE-POINT GEOSTATISTICS
HIERARCHICAL SIMULATION OF MULTIPLE-FACIES RESERVOIRS USING MULTIPLE-POINT GEOSTATISTICS AREPORT SUBMITTED TO THE DEPARTMENT OF PETROLEUM ENGINEERING OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE
More informationSelected Implementation Issues with Sequential Gaussian Simulation
Selected Implementation Issues with Sequential Gaussian Simulation Abstract Stefan Zanon (szanon@ualberta.ca) and Oy Leuangthong (oy@ualberta.ca) Department of Civil & Environmental Engineering University
More informationFeature Extractors. CS 188: Artificial Intelligence Fall Some (Vague) Biology. The Binary Perceptron. Binary Decision Rule.
CS 188: Artificial Intelligence Fall 2008 Lecture 24: Perceptrons II 11/24/2008 Dan Klein UC Berkeley Feature Extractors A feature extractor maps inputs to feature vectors Dear Sir. First, I must solicit
More informationMCMC Methods for data modeling
MCMC Methods for data modeling Kenneth Scerri Department of Automatic Control and Systems Engineering Introduction 1. Symposium on Data Modelling 2. Outline: a. Definition and uses of MCMC b. MCMC algorithms
More informationShort Note: Some Implementation Aspects of Multiple-Point Simulation
Short Note: Some Implementation Aspects of Multiple-Point Simulation Steven Lyster 1, Clayton V. Deutsch 1, and Julián M. Ortiz 2 1 Department of Civil & Environmental Engineering University of Alberta
More informationApplied Bayesian Nonparametrics 5. Spatial Models via Gaussian Processes, not MRFs Tutorial at CVPR 2012 Erik Sudderth Brown University
Applied Bayesian Nonparametrics 5. Spatial Models via Gaussian Processes, not MRFs Tutorial at CVPR 2012 Erik Sudderth Brown University NIPS 2008: E. Sudderth & M. Jordan, Shared Segmentation of Natural
More informationOur Calibrated Model has No Predictive Value: An Example from the Petroleum Industry
Our Calibrated Model as No Predictive Value: An Example from te Petroleum Industry J.N. Carter a, P.J. Ballester a, Z. Tavassoli a and P.R. King a a Department of Eart Sciences and Engineering, Imperial
More informationClassification: Feature Vectors
Classification: Feature Vectors Hello, Do you want free printr cartriges? Why pay more when you can get them ABSOLUTELY FREE! Just # free YOUR_NAME MISSPELLED FROM_FRIEND... : : : : 2 0 2 0 PIXEL 7,12
More informationThe exam is closed book, closed notes except your one-page cheat sheet.
CS 189 Fall 2015 Introduction to Machine Learning Final Please do not turn over the page before you are instructed to do so. You have 2 hours and 50 minutes. Please write your initials on the top-right
More informationComputer vision: models, learning and inference. Chapter 10 Graphical Models
Computer vision: models, learning and inference Chapter 10 Graphical Models Independence Two variables x 1 and x 2 are independent if their joint probability distribution factorizes as Pr(x 1, x 2 )=Pr(x
More informationPredictive Analysis: Evaluation and Experimentation. Heejun Kim
Predictive Analysis: Evaluation and Experimentation Heejun Kim June 19, 2018 Evaluation and Experimentation Evaluation Metrics Cross-Validation Significance Tests Evaluation Predictive analysis: training
More information