Bootstrap confidence intervals for reservoir model selection techniques

Size: px
Start display at page:

Download "Bootstrap confidence intervals for reservoir model selection techniques"

Transcription

1 Bootstrap confidence intervals for reservoir model selection techniques Céline Scheidt and Jef Caers Department of Energy Resources Engineering Stanford University Abstract Stochastic spatial simulation allows rapid generation of multiple, alternative realizations of spatial variables. Quantifying uncertainty on response resulting from those multiple realizations would require the evaluation of a transfer function on every realization. This is not possible in real applications, where one transfer function evaluation may be very time consuming (several hours to several days). One must therefore select a few representative realizations for transfer function evaluation and then derive the production statistics of interest (typically the P, P5 and P9 quantiles of the response). By selecting only a few realizations one may risk biasing the P, P5 and P9 estimates as compared to the original multiple realizations. The principle objective of this study is to develop a methodology to quantify confidence intervals for the estimated P, P5 and P9 quantiles when only a few models are retained for response evaluation. Our approach is to use the parametric bootstrap technique, which allows the evaluation of the variability of the statistics obtained from uncertainty quantification and construct confidence intervals. A second objective is to compare the confidence intervals when using two selection methods available to quantify uncertainty given a set of geostatistical realizations: traditional ranking technique and the distance-based kernel clustering technique (DKM). The DKM has been recently developed and has been shown to be effective in quantifying uncertainty. The methodology is demonstrated using two examples. The first example is a synthetic example, which uses bi-normal variables and serves to demonstrate the technique. The second example is from an oil field in West Africa where the uncertain variable is the cumulative oil production coming from 2 wells. The results show that for the same number of transfer function evaluations, the DKM method has equal or smaller error and confidence interval compared to ranking.

2 . Introduction Uncertainty quantification of subsurface spatial phenomena is done in the context of decision making, often by estimating low, mean and high quantile values (typically P, P5, and P9) of the response of interest. Often, an exhaustive sampling of all uncertain parameters is unfeasible, and only a small subset of reservoir model realizations of the phenomena can be created. Due to high computational requirements, the transfer function must be evaluated on an even smaller subset of realizations. Therefore, any quantiles that are estimated from this subset are themselves subject to uncertainty, and may vary depending on the selection method, the number of transfer function evaluations, the initial set of realizations, the use of a proxy response, etc. The objective of the study is to be able to quantify confidence intervals for the estimated P, P5 and P9 quantiles when only a few models are retained for response evaluation. The magnitude of the confidence intervals can then be used to decide whether or not more flow simulations are required to establish a better quantification of response uncertainty. The methodology developed uses parametric bootstrap technique, which is a statistical method allowing to construct confidence intervals of the estimated statistics. Such confidence intervals provide an idea on the variability of the statistics inferred by selecting only a few models for evaluation. The workflow can be applied using any technique of reservoir model selection. In this paper, we compare the behavior of the estimated quantiles using 3 different selection techniques. The first method is the traditional ranking technique (Ballin et al., 992), which select realizations according to a ranking measure. The second method has been developed recently and is called the distance-based kernel technique (DKM, Scheidt and Caers, 28). Finally, we use a random selection for comparison. It should be noted that the proposed bootstrap technique applies to any model selection methodology. The paper is organized as follows. In the next section, we give a description of the two methods employed to quantify uncertainty in spatial parameters. Then, we give a brief overview of the basic ideas of the bootstrap methodology in the context of parametric inference, illustrated by a typical example. We then describe our workflow which is applied to cases where we have a proxy response which can be evaluated rapidly for each realization, and a true response which cannot be evaluated for every realization. The subsequent section is devoted to the application of the specific workflow to two examples, the first being a synthetic example, the second is an example from an oil field in West Africa. Finally, we discuss the results obtained as well as some concluding remarks. 2. Quantification of uncertainty methodologies Uncertainty quantification of a spatial phenomenon aims at characterizing the 2

3 statistics (P, P5 and P9) of the response(s) of interest. In real applications where one transfer function evaluation can be very time consuming, it may not be possible to perform a transfer function evaluation on every realization of the reservoir. This difficulty can be overcome by selecting a representative set of realizations from the initial set. In this paper, we consider two different ways of selecting realizations for transfer function evaluation. The first method is the traditional ranking technique, which was introduced by Ballin et al. in 992. The second method, denoted the Distance-Kernel Method (DKM) is more recent and was first presented in 28 (Scheidt and Caers) and applied to a real case in Scheidt and Caers (29). 2.. Traditional Traditional ranking technique was introduced by Ballin in 992 in the context of stochastic reservoir modeling. The basic idea behind ranking is to define a rapidly calculable ranking measure, which can be evaluated for each realization. Most of the time, the ranking measure is static (eg. original oil-in-place), however more recent studies employ more complex measures, such as connectivity (McLennan and Deustch, 25), streamline (Gilman et al., 22) or tracer-based measures (Ballin et al., 992, Saad et al., 996). The ranking measure acts as a proxy of the response of interest for each realization. To be effective, therefore, ranking requires a good correlation between the ranking measure and the response. The ranking measure is used to rank the realizations according to the measure, and realizations are subsequently selected corresponding typically to the P, P5 and P9 quantiles. Full flow simulation is then performed on these selected realizations, and the P, P5 and P9 values are derived from the simulation results. In previous work (Scheidt and Caers, 29), we show that selecting only 3 realizations to derive the P, P5, and P9 quantiles can result in very inaccurate estimations. In this study, contrary to the standard ranking approach, we propose to select more than 3 realizations, and compare ranking with the Distance-Kernel Method proposed below. The realizations are selected equally-spaced according to the ranking measure, and we derive the P, P5 and P9 quantiles by interpolation from the distribution of the selected points Distance-Kernel Method In this section, we describe the main principle of the Distance-Kernel Method (DKM), illustrated in Figure. Starting from a large number of model realizations, the first step is to define a dissimilarity distance between the realizations. This distance is a measure of the dissimilarity between any two realizations, and should be tailored to the application and the response(s) of interest (just as in ranking), in order to make uncertainty quantification more efficient. The distance is evaluated between any two realizations, and a dissimilarity distance table ( R x R ) is then derived. Multi-dimensional scaling (MDS) is then applied using the distance table (Borg and 3

4 Groenen, 997). This results in a map (usually 2 or 3D) of the realizations, where the Euclidean distance between any two realizations is similar to the distance table. ote that only the distance between the realizations in the new space matters - the actual position of the realizations is irrelevant. Once the realizations are in MDS space, one could classify realizations and select a subset using clustering techniques. However, often the points in MDS space do not vary linearly and thus classical clustering methods would result in inaccurate classification. To overcome the nonlinear variation of the points, Schöelkopf et al. (22) introduced kernel methods to improve the clustering results. The main idea behind kernel methods is to introduce a highly non-linear function Φ and map the realizations from the MDS space to a new space, called feature space. The high dimensionality of that space makes the points behave more linearly and thus standard classification tools, such as clustering, can be applied more successfully. In this paper, we employ kernel k-means to select representative realizations of the entire set. Transfer function evaluation is then applied on the closest realization to the centroids and the statistics (P, P5 and P9) are computed on the small subset of realizations. Model Model 2 δ 2 δ 3 δ 4 δ 24 δ 32 δ 34 Model 3 Model Distance Matrix D δ δ 2 δ 3 δ 4 2 δ 2 δ 22 δ 32 δ 42 Euclidean Space on-linear variation (a) (b) (c) 3 δ 3 δ 23 δ 33 δ 43 4 δ 4 δ 24 δ 34 δ 44 R simulation P,P5,P9 model selection Kernels Φ linear features (e) R Φ - Apply Apply standard standard tools tools here here PCA PCA Clustering Clustering Dimensionality Dimensionality reduction reduction (d) F Figure : DKM for uncertainty quantification: (a) distance between two models, (b) distance matrix, (c) models mapped in Euclidean space, (d) feature space, (e) preimage construction, (f) P, P5, P9 estimation 4

5 For more details about the methodology, we refer to Scheidt and Caers (28). 3. Parametric Bootstrap Methodology 3.. General introduction to Bootstrap Bootstrap methods fall within the broader category of resampling methods. The concept of the bootstrap was first introduced by Efron (979). In his paper, Efron considered two types of bootstrap procedures (nonparametric and parametric inference). Bootstrap is a Monte-Carlo simulation technique that uses sampling theory to estimate the standard error and the distribution of a statistic. In many recent statistical texts, bootstrap methods are recommended for estimating sampling distributions, finding standard errors and confidence intervals. A bootstrap procedure is the practice of estimating properties of an estimator (such as its variance) by measuring those properties when sampling from an approximate distribution. In the parametric bootstrap, we consider an unknown distribution F to be a member of some prescribed parametric family and obtain the empirical distribution Fˆ n by estimating the parameters of the family from the data. Then, a new random sequence, called a resample, is generated from the distribution Fˆ n. The parametric bootstrap procedure works as follows. First, the statistics θˆ of the distribution of the initial sample are computed (for example the mean and variance). Then, the distribution Fˆ n is estimated using those statistics. We assume that the distribution Fˆ n is the true distribution and we use Monte-Carlo simulation to generate B new samples of the initial sample using the distribution Fˆ n. ext, we apply the same estimation technique to these bootstrapped data to get a total of B bootstrap estimates of θˆ, which are denoted ˆ*b θ, b =, B. Using these B bootstrap estimates, we can compute confidence intervals or any other statistical measure of error. Simple illustrative example A simple example illustrating the parametric bootstrap method is presented in Figure 2. Suppose we have R =5 values X = (x,,x R ) of a normal distribution ( µ, σ ) and we are interested in the estimation of the unknown parameters µ and σ. The first step is to assume that X has a normal distribution F n and then to estimate the mean and variance of the distribution: µˆ = x and ( ) 2 R 2 ˆ σ = x i x R i= 5

6 We assume that µˆ and σˆ are the true parameters and we generate B = new samples X (b=,,b) from Fˆ n = ( ˆ, µ ˆ σ ) using Monte-Carlo simulation, each sample containing R =5 values. For each sample, the bootstrap estimate of the mean and variance of the distribution can be calculated: ˆ = µ b x * Having computed ˆ b 2* b ( ) * ˆ µ, ˆ σ R 2 and ˆ σ = ( xi x )2 R i= θ =, one can now construct a histogram of the mean and the variance to display the probability distribution of the bootstrap estimator (Figure 2). From this distribution, one can obtain an idea of the statistical properties of the estimates µˆ and σˆ. In Figure 2, the red line represents the estimation of the mean µˆ and variance σˆ of the initial sample. F ( µ, σ ) X = [ x,..., x ] R ˆ θ = ( ˆ, µ ˆ) σ Fˆ ( ˆ, µ ˆ) σ Frequency 2 5 Frequency X = [ x,..., x R ] Bootstrap Mean Bootstrap Variance ˆ * b θ = ( ˆ µ, ˆ σ ) Figure 2: Application of the parametric bootstrap procedure to a simple example The histograms of the bootstrap estimations of the mean and the variance are informative about the variability of the statistics obtained. Confidence intervals of the estimated mean and variance (or any quantiles) can then be calculated from the B estimates of the mean and variance. The next section shows how to apply the bootstrap method in the context of uncertainty quantification where a proxy value can be rapidly calculated for many realizations of a spatial phenomenon Workflow of the study Contrary to the previous example where the data are univariate, in the context of reservoir model selection techniques, a proxy response is employed for the selection 6

7 using DKM or ranking and thus two variables are necessary: the response of interest and the proxy response. Therefore, we consider a bivariate variable X = [X, X 2,, X R ], where: X i = [x i,y i ], i =,, R, R being the total number of samples/realizations x i represents the response of interest (e.g. cumulative oil production) y i represents the proxy response, which will serve as a ranking measure or be transformed to a distance. ote that for ranking and DKM to be effective, the response and its proxy should be reasonably well correlated. In addition, for real applications, the values of the true response x i for each realization are unknown. In traditional uncertainty quantification, the proxy response serves as a guide to select a few realizations which will be evaluated using the transfer function. The response quantiles are then deduced from the evaluation of the realizations. Since the resulting quantiles are subject to uncertainty, the bootstrap method illustrated previously is well suited to the problem and can inform us on the accuracy of the estimated quantiles and give an idea of the error resulting from the selection of a small subset of realizations. The workflow in the context of reservoir model selection is as follow. It is illustrated in Figure 3.. Evaluate a proxy response y i for each of the i =,, R realizations. 2. Apply ranking or DKM using the proxy response in order to select samples/realizations for evaluation (note that << R ). Compute the statistics of interest - P, P5 and P9 in the case of uncertainty ˆ * * * θ = xˆ, xˆ, xˆ. quantification: ( ) P P5 P9 3. Assume that the distribution of X is a member of some parametric family and fit a bivariate parametric model Fˆ n by estimating the family parameters from the selected data. 4. Assume that Fˆ n is the true distribution and use Monte-Carlo simulation to generate B new samples from this parametric model Fˆ n. For each of the B samples generated, apply ranking or DKM to select realizations and compute the statistics of interest: ˆ * * * = ( xˆ, ˆ, ˆ P xp xp ) θ. 5. From the B samples of θˆ, compute the confidence intervals on any statistics of interest. One way to estimate confidence intervals from bootstrap samples is to take the α and -α quantiles of the estimated values (α = in this study). These are called bootstrap percentile 5 9 7

8 intervals. 6. A single measure of accuracy of our quantile estimation is defined by computing the dimensionless bootstrap error of the estimated quantiles for each of the B new samples created (Eq. ): xˆ P xˆ P xˆ P xˆ ˆ ˆ 5 P x 5 P x 9 P 9 error = () xˆ ˆ ˆ P x P x 5 P9 The bootstrap error of the estimated quantiles is evaluated on each sample, and thus can be represented as a histogram to visualize the variability between the samples. From the histogram, we can quantify the variation of the bootstrap error of the estimated quantiles. True Values x, K, x R y, K y, R / -> Select points x K x * *,, y K y * *,, Parametric Bootstrap Estimation of distribution Fˆn Generation of B samples xˆ * P, xˆ * P 5, xˆ * P 9 b =,,B x,k, y,k, x y, K, x ( b ) ( b ) / R -> Select points ( b ) ( b ) y, K, y x R xˆ P, xˆ P 5, xˆ P 9 error = 3 xˆ P xˆ xˆ P P + xˆ P5 xˆ xˆ P5 P5 + xˆ P9 xˆ xˆ P9 P9 Figure 3: Workflow of the bootstrap method applied to uncertainty quantification The workflow described previously and illustrated in Figure 3 is performed for several values of, where is the number of selected realizations for evaluation. This is done to evaluate the influence of the number of transfer function evaluations on the accuracy of the quantile estimation. For each value of, the selected realizations are obtained using DKM or ranking methods, and therefore the realizations are different for each value of. ow that the basic idea and theory of the bootstrap method has been presented, 8

9 the next section shows some application examples of this technique in the context of uncertainty quantification. 4. Application of the methodology to uncertainty quantification Two examples are presented in this section. The first one is illustrative and uses a bivariate Gaussian distribution. The second example is more complex and is based on a real oil field reservoir in West Africa (West Coast African reservoir) and uses real production data. In the case of DKM, the definition of a distance between any two realizations is required. In this study, in order to compare the results of the DKM with those obtained by ranking using the exact same information, we use simply the difference of ranking measure (proxy response) as a distance between realizations. ote however that, as opposed to the ranking measure, the distance can be calculated using a combination of many different measures, and thus has more flexibility to be tailored to the application. We will discuss the consequences of this in more detail below. 4.. Bivariate Gaussian distribution In the first example, we consider a bivariate Gaussian distribution: X ~ bi ( µ, Σ), where µ represents the mean and Σ the covariance matrix. In this example, the mean of the sample is taken as µ = [5,5], and the covariance is taken 2 2ρ as: Σ =. The parameter ρ defines the correlation coefficient between the 2ρ 2 target response and the proxy response. To set up an example, an initial sample X of R = values is generated using the distribution given above. ote that for this illustrative example, we use the term sample instead of realization, since no geostatistical realization is associated to each bivariate value. Figure 4 shows an example of the probability density plot of the binormal sample X, where the correlation coefficient between the target and proxy responses was defined as ρ =.9. 9

10 Figure 4: Probability density of X, which has a bi-normal distribution ow that the initial data is defined, we assume that we only know the type of distribution of X (bi-normal), but that we do not know the parameters defining the distribution (the mean µ and the covariance Σ ). The bootstrap procedure illustrated in Figure 3 is applied taking the sample X generated previously (Figure 4) and using DKM to select =5 points. Estimation of the mean µˆ and the covariance Σˆ are then obtained using the response on the 5 selected points and thus the estimated bivariate distribution of the data is assumed to be the true distribution: Fˆ n = bi( ˆ, µ Σˆ ). B= new samples of this distribution can then be easily derived, since the distribution is known. Uncertainty quantification is then performed on those B samples, and an estimation of the variability of the quantiles is possible. Examples of the bootstrap histograms of the P, P5 and P9 quantiles are presented in Figure Frequency Frequency Frequency Bootstrap P Bootstrap P Bootstrap P9 x ˆ * P Figure 5: Histogram of the P, P5 and P9 quantiles estimated by bootstrap, xˆ * P5, xˆ * P9. The red line represents the estimated quantiles x ˆ * P, xˆ estimates are calculated using DKM to select 5 points. * P5, xˆ * P9. The

11 We observe on Figure 5 that the distribution of the bootstrap quantiles is Gaussian. In addition, there is a small bias in the estimation of the P and P9 quantiles for this example. Although this is not shown, ranking has the same effect. The result is that on average, the xˆp is overestimated and the x ˆ P9 is underestimated. The biased estimates should not affect the determination of the confidence intervals. In our study, we have found that the estimated mean µˆ and covariance Σˆ from the initial sample had an impact on the confidence intervals. Since our goal in this first example is to understand what the general behavior is when varying the number of selected samples, we propose to do a Monte-Carlo bootstrap, which basically means that we randomize the initial sample and use C sets of initial samples, then perform the workflow illustrated in Figure 3 on those C sets of initial samples. The estimated statistics of each initial sample are averaged to obtain the final statistics. In this study, we take C = 5. In the next few examples, the workflow illustrated in Figure 3 has been performed by varying the number of selected samples ( = 5, 8,, 5 and 2 more precisely), in order to examine the effect of the number of transfer function evaluations on the bootstrap error. In addition, several correlation values between the proxy response and the target response were used to explore the influence of the correlation coefficient on the confidence intervals. Results are presented in Figure 6, for ρ =,.9,.8,.7,.6 and.5 respectively. Figure 6 shows the confidence intervals of the error of the bootstrap estimated quantiles for DKM (blue - square) and ranking (red - dot) for different values of. The number of bootstrap samples generated is B =. The symbols represent the P5 value of the error of the estimated quantiles, in other words, half of the estimated quantiles have an error below this value and half above.

12 ρ =. ρ =.9 Error on quantile estimation Error on quantile estimation ρ =.8 ρ =.7 Error on quantile estimation Error on quantile estimation ρ =.6 ρ =.5 Error on quantile estimation Error on quantile estimation Figure 6: Confidence intervals (α = ) of the bootstrap error of the estimated quantiles as a function of the number of function evaluation for ρ =,.9,.8,.7,.6 and.5. The symbols represent the P5 value of the bootstrap error. We observe on Figure 6 that the error globally decreases as the number of transfer function evaluation increases. Also, the confidence intervals tend to narrow as the number of transfer function evaluation increases, meaning that the error in our estimates decreases. Both methods, DKM and ranking, provide similar results. However, the error obtained by the DKM is slightly smaller than the one observed for ranking. The same remark is valid for the confidence intervals. Finally, the results 2

13 provided by DKM vary smoother than the one obtained by ranking technique. ote that each method selects optimally samples for evaluation. Therefore, the = 8 models selected do not necessarily include the = 5 models. This is true for all. The bootstrap method can also be used to compute an estimate of the correlation coefficient between the actual response and the proxy response. Figure 7 presents the confidence intervals for the correlation corresponding to the results obtained in Fig. 6. ρ =.9 ρ =.8.95 Estimated correlation Estimated correlation ρ =.7.2 ρ =.6.9 Estimated correlation Estimated correlation ρ =.5 Estimated correlation Figure 7: Bootstrap estimates (α = ) of the correlation between the response and the proxy. The black line represents the input correlation (ρ =.9,.8,.7,.6 and.5) to generate the first sample 3

14 We observe on Figure 7 that the estimation of the correlation coefficient tends to be overestimated, especially for small values of. Figure 7 also shows that the correlation estimation becomes more accurate and less prone to error as the number of transfer function evaluation increases. The next section illustrates the workflow using a real oil reservoir, located in West Africa West Coast African reservoir Reservoir Description The West Coast African (WCA) reservoir is a deepwater turbidite offshore reservoir located in a slope valley. The reservoir is located offshore in 6 feet of water and is 46 feet below see level. Four depositional facies were interpreted from the well logs: shale (Facies ), poor quality sand # (Facies 2), poor quality sand #2 (Facies 3) and good quality channels (Facies 4). The description of the facies filling the slope valley is subject to uncertainty. 2 TIs are used in this case study, representing uncertainty on the facies representations. The reservoir is produced with 28 wells, of which 2 are production wells and 8 are water injection wells. The locations of the wells are displayed in Figure 8. Wells colored in red are producers wells and in blue are injectors. Figure 8: Location of the 28 wells. Red are production wells and blue are injection wells. Different colors in grid represent different fluid regions 72 geostatistical realizations were created using the multi-point geostatistical algorithm snesim (Strebelle, 22). To include spatial uncertainty, two realizations were generated for each combination of TI and facies probability cube, leading to a total of 72 possible realizations of the WCA reservoir. Each flow simulation took 4.5 hours. 4

15 In a previous paper (Scheidt and Caers, 29), uncertainty quantification on the WCA reservoir has been performed by performing only a small number of simulations. It was shown that the statistics obtained by flow simulation on 7 realizations selected by the DKM are very similar to the one obtained by simulation on the entire set of 72 realizations. A comparison with the traditional ranking method showed that the DKM method easily outperforms ranking technique without requiring any additional information. However, in reality, one would not have access to the results of those 72 flow simulations, hence one would not know how accurate the results of P, P5, P9 of those 7 flow simulations are with respect to the entire set of 72 flow simulations. In this study, the response of interest is the cumulative oil production at 2 days. We have evaluated the response for each of the 72 realizations, as a reference. For the proxy response, we evaluated the cumulative oil production using streamline simulation (Batycky et al., 997). The correlation coefficient between the response and the proxy is ρ =.92. In order to perform the parametric bootstrap procedure, we must estimate the distribution of the cumulative oil production and its ranking proxy, and be able to generate new samples with densities of the bivariate distribution. Because we do not know a priori the distribution of the cumulative oil production and its proxy (contrary to the previous example), we propose to compute the bivariate densities using a kernel smoothing technique (Bowman and Azzalini, 997). Generation of a sample for a kernel smoothing density Kernel smoothing (Bowman and Azzalini, 997) is a spatial method that generates a map of density values. The density at each location reflects the concentration of points in the surrounding area. Kernel smoothing does not require making any parametric assumption about the probability density function (pdf). The kernel smoothing density of a variable X = x, K, ] is defined as follow: [ x R R x xi fˆ( x, h) = K p h i= h with K the kernel function and h the bandwidth. R, x, x i R In the case of a Gaussian rbf kernel, the kernel smoothing density is defined as: 2 R x xi fˆ( x, h) = exp p / 2 Rh (2π ) i= 2 h Choosing the bandwidth for the kernel smoothing can be a difficult task, and is generally a compromise between acceptable smoothness of the curve and fidelity to the data. The choice of h has an impact on the overall appearance of the resulting smooth curve, much more so than the choice of the kernel function which is generally held to be of secondary importance. In this work, we use a bandwidth which is function to the number of points in X. p 5

16 For example, Figure 9 shows the density distribution of the 72 data from the WCA example, estimated by kernel smoothing using a Gaussian kernel. Figure 9: Probability density of X for WCA Once the density of the bivariate variable has been defined, new samples of the same distribution can be generated using Metropolis sampling technique. Overview of the Metropolis sampling algorithm Metropolis-Hasting technique is a Markov chain-based method which allows generating a random variable having a particular distribution (Metropolis and Ulam 949, Metropolis et al. 953). The Metropolis algorithm generates a sequence of samples from a distribution f as follows:. Start with some initial value x 2. Given this initial value, draw a candidate value x* from some proposal distribution (we choose a uniform distribution) 3. Compute the ratio α of the density at the candidate x* and the current x t- points and accept the candidate point with probability α: α = f ( x*) f ( ) x t 4. Return to step 2 until the desired number of samples is obtained. 5. The new sample ( x,..., ) x t is of distribution f 6

17 An illustration of a sample generated by Metropolis sampling associated with the density provided by kernel smoothing is presented in Figure. The contours present the density probability which is calculated using = values of response selected by DKM. The red points show 7 values derived from this density by Metropolis sampling. Figure : Generation of a new sample using Metropolis sampling. The contours represent the probability density obtained by kernel smoothing and red dots represent new sample generated by Metropolis sampling Application of the bootstrap technique to the WCA case In the context of uncertainty quantification in cumulative oil production, the initial data are the flow simulations at the R = 72 realizations of the WCA reservoir: x, x R : Cumulative oil production obtained by full flow simulation (target response) y, y R : Cumulative oil production obtained by fast flow simulation (proxy response) The distance employed for the DKM is computed as the absolute value of the difference of proxy response between any two realizations: d ij = y y. The bootstrap procedure, illustrated in Figure 3, is performed for different number of transfer function evaluation: in this case = 3, 5, 8,, 5 and 2. For each value of, the procedure to generate B bootstrap samples is as follow:. Select realizations using the proxy response as ranking measure or distance measure d according to the method used 2. Evaluation of the response using the transfer function (flow simulation) on the selected realizations i j 7

18 3. Compute the bivariate density Fˆ n of the target response using kernel smoothing on the responses resulting from the selected realizations 4. Use Metropolis sampling to generate B samples of the bivariate distribution Fˆ n 5. For each of the B samples generated, apply ranking or DKM to select realizations and compute the statistics of interest: ˆ * * * = ( xˆ, ˆ, ˆ P xp xp ) θ. The workflow illustrated in Figure gives more details than the general workflow in Figure 3, by including the estimation of Fˆ n by kernel smoothing and the generation of new samples by metropolis sampling. 5 9 Proxy measure: y, K y, R /ranking to select real. Response evaluation on selected real. Kernel Smoothing on selected realizations Metropolis sampling to generate a new sample x K x y K y * *,, * *,, Fˆn x y, K, x ( b ) ( b ) R, K, y ( b ) ( b ) R /ranking to select real. P, P5 and P9 evaluation on selected real. x,k, y,k, x y Figure : Workflow for confidence interval calculation The next section shows an application of the workflow illustrated above in Figure. The workflow is performed using 3 different methods for selecting the realizations: DKM, ranking and random selection. Our objective is to see how each method behaves as the number of transfer function evaluation increases and how they compare to each other. First, we compare the 3 methods by looking at the histograms of the bootstrap error of the estimated quantiles for each method (Figure 2). The bootstrap error is computed using Eq. above. The blue, red and green bars represent the error obtained for DKM, ranking and random selection respectively. 8

19 3 25 = 3 DKM Random 3 25 = 5 DKM Random Frequency 2 5 Frequency Frequency Response Value = 8 DKM Random Response Value Frequency Response Value = DKM Random Response Value 3 25 = 5 DKM Random 3 25 = 2 DKM Random Frequency 2 5 Frequency Response Value Response Value Figure 2: Histograms of the bootstrap error of the estimated quantiles for different number of function evaluation and 3 selection methods. We observe that, in each case, the DKM method performs better than the ranking technique. For all values of, the errors are globally smaller for the DKM than for ranking or random selection. In addition, the error variance is reduced with more transfer function evaluations. 9

20 Figure 2 represents the bootstrap percentile intervals (α = ) of the bootstrap error of the estimated quantiles. The symbol in each interval represents the P5 value of the error. Error on quantile estimation Quantiles estimation Random Figure 3: Confidence intervals (α = ) of the bootstrap error of the estimated quantiles as a function of the number of function evaluations We observe on Figure 3 that the error tends to decrease as the number of function evaluations increases. As observed before on the histograms, DKM performs better than ranking, which performs better than random selection. This conclusion was also reached in Scheidt and Caers (29). In this example, we observe that for the DKM, the results stabilize for > 8. We can therefore conclude that 8 or flow simulations are necessary for the DKM selected models to have the same uncertainty as the total set of 72. In a previous paper (Scheidt and Caers, 29), it was concluded that 7 simulations were satisfactory. ote however that the distance in that work was slightly more correlated to the difference in response compared to the correlation in this study. The table below represents the mean of the bootstrap error, computed from the histograms presented in Figure 2. DKM Random select. = = = = = = Table : Mean of the dimensionless bootstrap error for each selection method. 2

21 This table, as well as the histograms and confidence intervals can be very useful to give an indication of the error resulting from the quantile estimation of the response, based on the selected realizations. For example, suppose we are limited in time and can only perform 5 transfer function evaluations. However, we want to be sure that we can be confident on the uncertainty quantification results derived from those 5 simulations. From Table, we can see that the mean error for = 5 for DKM is.333 and.495 for ranking. If we had a little more time and had performed = 8 simulations, the error would be.28 and.293, which is an improvement of 6% (5.8% for ranking) compared to the results from = 5. Another way of looking at the results is to show the confidence intervals for each quantile individually. This is illustrated on Figure x 4 Quantiles estimation 7 Random 7.6 x Quantiles estimation P 6.5 P x Quantiles estimation Random P Random Figure 4: Confidence intervals of the bootstrap estimates of the quantiles P, P5 and P9 (BBL) as a function of the number of function evaluation. The line represents the quantiles derived from the entire set of realizations Figure 4 shows that DKM and ranking produce very accurate estimates of the P5 quantile of the target response, even for small number of transfer function evaluations (medians are easier to estimate than extremes). In addition, the P quantiles tend to be slightly underestimated, but DKM is closest to the reference 2

22 value than the other techniques. The same conclusions are valid for the P9, except that we observe an overestimation of the quantiles. The underestimation of P and overestimation of P9 are most likely due to the use of kernel smoothing to estimate the density, which will increase the variability of the response compared to the original 72 realizations. As mentioned in the beginning of the paper, the proxy measure should be correlated for DKM and ranking to be effective. However, the correlation coefficient between both responses is not known a priori, since the target response for all realizations is unknown. Once a selection method is applied and the transfer function is evaluated on the selected realizations, an estimation of the correlation coefficient can be inferred. The quality of the estimated correlation coefficient can be studied in exactly the same way than the estimated quantiles, by doing parametric bootstrap. Figure 5 represents the confidence intervals obtained for different values of, the correlations being estimated on the same samples used to estimate the quantile error. The symbols show the initial estimates of the covariance ρˆ. Quantiles estimation Estimated correlation coefficient Figure 5: Bootstrap estimated correlation coefficient on the WCA test case. Figure 5 shows that the first estimates ρˆ of the correlation coefficient between the 2 responses are accurate, with a slight overestimation for small number of transfer function evaluations ( = 3 and 5). However, the bootstrap estimated correlation coefficients are largely underestimated. We believe that this is due to the kernel smoothing technique, which tends to smooth the density of the bivariate data and therefore allow Metropolis sampling to sample points in a wider area than it should. This was not the case for the bi-normal example in Section 4.. However, we can still derive conclusions on the confidence intervals provided. We observe that DKM tends to have less uncertainty in the correlation coefficient than ranking, except for = 8. 22

23 5. Discussion on distances The above examples demonstrate that using the same measure for ranking and distance provides for similar accuracy in uncertainty quantification for the Gaussian case. We should emphasize however that the bootstrap method applied in the context of the paper is clearly unfavorable to DKM. In order to compare ranking and the DKM, we calculated the distance between 2 realizations as the difference of the ranking measure between the realizations. This leads to a representation of uncertainty in a D MDS-space, and therefore the use of kernel methods has not the same impact as for higher dimensional MDS-space. The distance in this study is very simple, whereas in many applications the distance can be much more complex, and can take into account many measures of influential factors on the response. For example, a distance can be a function of many parameters, such as the cumulative oil production at different times, and water-cut of a a group of wells (Scheidt and Caers, 29). Using traditional ranking techniques may require multiple independent studies if one is interested in uncertainty in several responses. In the case of DKM, a single study is enough if the distance is well chosen. 6. Conclusions We have established a new workflow to construct confidence intervals on quantile estimations in model selection techniques. We would like to state explicitly that we do not treat the question of whether or not the uncertainty model, i.e. the possibly large set of reservoir models that can be generated by varying several input parameters, is realistic. Uncertainty quantification by itself is inherently subjective and any confidence estimates of the uncertainty model itself are therefore useless. In this paper we assume there is a larger set of model realizations and assume that this set provides a realistic representation of uncertainty. Then, the proposed bootstrap allows quantifying error on uncertainty intervals or quantiles when only a few models from the larger set are selected. The workflow uses model selection methods in this work DKM or ranking - and employs a parametric bootstrap procedure to construct confidence intervals on the quantiles retained by the model selection techniques. Examples show that DKM provides more robust results compared to ranking, especially for small number of transfer function evaluations. The study of the uncertainty resulting from model selection can be very informative - it shows if we can be confident or not in the estimated statistics. The confidence interval is a function of the estimated variance of the response and the estimated correlation coefficient between the proxy measure and the response. Since the user does not know the correlation coefficient a priori, we propose performing a bootstrap procedure between the response and its proxy to estimate the quality of the distance. If the estimated correlation coefficient is high and its associated uncertainty low, then we can be confident on the uncertainty quantification results. If after transfer function evaluations the uncertainty is large 23

24 and a poor correlation is found, then the results should be improved by either using a better proxy response or doing more transfer function evaluations. omenclature R = number of initial realizations = number of selected realizations for transfer function evaluation X = [X,, X R ] Xi = [x i, y i ] x i = target response value for realization i y i = proxy response value for realization i d ij = distance between realizations i and j ρ = correlation coefficient between the target and proxy responses B = number of samples generated in the bootstrap procedure e = bootstrap error of the estimated quantiles for sample b * * * x ˆ, xˆ, xˆ = estimated P, P5 and P9 after the first selection method P * P P5 * P5 P9 x ˆ, xˆ, xˆ = bootstrap estimated quantiles for the second selection method * P9 References Ballin, P.R., Journel A.G., and Aziz, K. [992] Prediction of Uncertainty in Reservoir Performance Forecast, JCPT, no. 4. Batycky, R. P., Blunt, M. J. and Thiele, M. R A 3D Field-Scale Streamline-Based Reservoir Simulator, SPERE 2(4): Borg, I., Groenen, P Modern multidimensional scaling: theory and applications. ew-york, Springer.Bowman, A. W., and A. Azzalini, [997] Applied Smoothing Techniques for Data Analysis, Oxford University Press Erfon, B. [979]. Bootstrap methods: Another look at the Jackknife, The Annals of Statistics 7 (): -26 Hastings, W. K. 97. Monte Carlo sampling methods using Markov Chains and their applications. Biometrika 57: McLennan, J.A., and Deutsch, C.V. 25. Geostatistical Realizations by Measures of Connectivity, Paper SPE/PS-CIM/CHOA 9868 presented at the SPE International Thermal Operations and Heavy Oil Symposium, Calgary, Alberta, Canada, -3 ovember. Metropolis,., and S. Ulam The Monte Carlo method. J. Amer. Statist. Assoc. 44: Metropolis,., A.W. Rosenbluth, M.. Rosenbluth, A.Teller, and H. Teller. 24

25 953. Equations of state calculations by fast computing machines. Journal of Chemical Physics 2: 87 9 Saad,., Maroongroge, V. and Kalkomey C. T Geostatistical Models Using Tracer Production Data, Paper presented at the European 3-D Reservoir Modeling Conference, Stavanger, orway, 6-7 April. Scheidt, C., and Caers, J. 28. Representing Spatial Uncertainty Using Distances and Kernels. Mathematical Geosciences, DOI:.7/s Scheidt, C., and Caers, J. 29, A new method for uncertainty quantification using distances and kernel methods. Application to a deepwater turbidite reservoir. Accepted in SPEJ. To be published. Schoelkopf B., Smola A. (22) Learning with kernels, MIT Press, Cambridge, 664p. Strebelle, S. 22. Conditional Simulation of Complex Geological Structures using Multiple-point Statistics, Mathematical Geology, 34():

We LHR3 04 Realistic Uncertainty Quantification in Geostatistical Seismic Reservoir Characterization

We LHR3 04 Realistic Uncertainty Quantification in Geostatistical Seismic Reservoir Characterization We LHR3 04 Realistic Uncertainty Quantification in Geostatistical Seismic Reservoir Characterization A. Moradi Tehrani* (CGG), A. Stallone (Roma Tre University), R. Bornard (CGG) & S. Boudon (CGG) SUMMARY

More information

Advanced analysis and modelling tools for spatial environmental data. Case study: indoor radon data in Switzerland

Advanced analysis and modelling tools for spatial environmental data. Case study: indoor radon data in Switzerland EnviroInfo 2004 (Geneva) Sh@ring EnviroInfo 2004 Advanced analysis and modelling tools for spatial environmental data. Case study: indoor radon data in Switzerland Mikhail Kanevski 1, Michel Maignan 1

More information

Statistical Rock Physics

Statistical Rock Physics Statistical - Introduction Book review 3.1-3.3 Min Sun March. 13, 2009 Outline. What is Statistical. Why we need Statistical. How Statistical works Statistical Rock physics Information theory Statistics

More information

Assessing uncertainty on Net-to-gross at the Appraisal Stage: Application to a West Africa Deep-Water Reservoir

Assessing uncertainty on Net-to-gross at the Appraisal Stage: Application to a West Africa Deep-Water Reservoir Assessing uncertainty on Net-to-gross at the Appraisal Stage: Application to a West Africa Deep-Water Reservoir Amisha Maharaja April 25, 2006 Abstract A large data set is available from a deep-water reservoir

More information

Reservoir connectivity uncertainty from stochastic seismic inversion Rémi Moyen* and Philippe M. Doyen (CGGVeritas)

Reservoir connectivity uncertainty from stochastic seismic inversion Rémi Moyen* and Philippe M. Doyen (CGGVeritas) Rémi Moyen* and Philippe M. Doyen (CGGVeritas) Summary Static reservoir connectivity analysis is sometimes based on 3D facies or geobody models defined by combining well data and inverted seismic impedances.

More information

Best Practice Reservoir Characterization for the Alberta Oil Sands

Best Practice Reservoir Characterization for the Alberta Oil Sands Best Practice Reservoir Characterization for the Alberta Oil Sands Jason A. McLennan and Clayton V. Deutsch Centre for Computational Geostatistics (CCG) Department of Civil and Environmental Engineering

More information

Modeling uncertainty in metric space. Jef Caers Stanford University Stanford, California, USA

Modeling uncertainty in metric space. Jef Caers Stanford University Stanford, California, USA Modeling uncertainty in metric space Jef Caers Stanford University Stanford, California, USA Contributors Celine Scheidt (Research Associate) Kwangwon Park (PhD student) Motivation Modeling uncertainty

More information

Reliability of Seismic Data for Hydrocarbon Reservoir Characterization

Reliability of Seismic Data for Hydrocarbon Reservoir Characterization Reliability of Seismic Data for Hydrocarbon Reservoir Characterization Geetartha Dutta (gdutta@stanford.edu) December 10, 2015 Abstract Seismic data helps in better characterization of hydrocarbon reservoirs.

More information

A MultiGaussian Approach to Assess Block Grade Uncertainty

A MultiGaussian Approach to Assess Block Grade Uncertainty A MultiGaussian Approach to Assess Block Grade Uncertainty Julián M. Ortiz 1, Oy Leuangthong 2, and Clayton V. Deutsch 2 1 Department of Mining Engineering, University of Chile 2 Department of Civil &

More information

A008 THE PROBABILITY PERTURBATION METHOD AN ALTERNATIVE TO A TRADITIONAL BAYESIAN APPROACH FOR SOLVING INVERSE PROBLEMS

A008 THE PROBABILITY PERTURBATION METHOD AN ALTERNATIVE TO A TRADITIONAL BAYESIAN APPROACH FOR SOLVING INVERSE PROBLEMS A008 THE PROAILITY PERTURATION METHOD AN ALTERNATIVE TO A TRADITIONAL AYESIAN APPROAH FOR SOLVING INVERSE PROLEMS Jef AERS Stanford University, Petroleum Engineering, Stanford A 94305-2220 USA Abstract

More information

Optimizing Thresholds in Truncated Pluri-Gaussian Simulation

Optimizing Thresholds in Truncated Pluri-Gaussian Simulation Optimizing Thresholds in Truncated Pluri-Gaussian Simulation Samaneh Sadeghi and Jeff B. Boisvert Truncated pluri-gaussian simulation (TPGS) is an extension of truncated Gaussian simulation. This method

More information

COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION

COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION G. MARIETHOZ, PH. RENARD, R. FROIDEVAUX 2. CHYN, University of Neuchâtel, rue Emile Argand, CH - 2009 Neuchâtel, Switzerland 2 FSS Consultants, 9,

More information

Reservoir Uncertainty Calculation by Large Scale Modeling

Reservoir Uncertainty Calculation by Large Scale Modeling Reservoir Uncertainty Calculation by Large Scale Modeling Naeem Alshehri and Clayton V. Deutsch It is important to have a good estimate of the amount of oil or gas in a reservoir. The uncertainty in reserve

More information

Geostatistical Determination of Production Uncertainty: Application to Firebag Project

Geostatistical Determination of Production Uncertainty: Application to Firebag Project Geostatistical Determination of Production Uncertainty: Application to Firebag Project Abstract C. V. Deutsch, University of Alberta (cdeutsch@civil.ualberta.ca) E. Dembicki and K.C. Yeung, Suncor Energy

More information

Multiple-Point Geostatistics: from Theory to Practice Sebastien Strebelle 1

Multiple-Point Geostatistics: from Theory to Practice Sebastien Strebelle 1 Multiple-Point Geostatistics: from Theory to Practice Sebastien Strebelle 1 Abstract The limitations of variogram-based simulation programs to model complex, yet fairly common, geological elements, e.g.

More information

Direct forecasting without full model inversion Jef Caers

Direct forecasting without full model inversion Jef Caers Direct forecasting without full model inversion Jef Caers Professor of Geological Sciences Stanford University, USA Why moving to Geological Sciences? To attract a greater diversity of students To be able

More information

Inverting hydraulic heads in an alluvial aquifer constrained with ERT data through MPS and PPM: a case study

Inverting hydraulic heads in an alluvial aquifer constrained with ERT data through MPS and PPM: a case study Inverting hydraulic heads in an alluvial aquifer constrained with ERT data through MPS and PPM: a case study Hermans T. 1, Scheidt C. 2, Caers J. 2, Nguyen F. 1 1 University of Liege, Applied Geophysics

More information

COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION

COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION COLLOCATED CO-SIMULATION USING PROBABILITY AGGREGATION G. MARIETHOZ, PH. RENARD, R. FROIDEVAUX 2. CHYN, University of Neuchâtel, rue Emile Argand, CH - 2009 Neuchâtel, Switzerland 2 FSS Consultants, 9,

More information

Analysis of the Pattern Correlation between Time Lapse Seismic Amplitudes and Saturation

Analysis of the Pattern Correlation between Time Lapse Seismic Amplitudes and Saturation Analysis of the Pattern Correlation between Time Lapse Seismic Amplitudes and Saturation Darkhan Kuralkhanov and Tapan Mukerji Department of Energy Resources Engineering Stanford University Abstract The

More information

Quantifying asymmetric parameter interaction. in sensitivity analysis: application to reservoir

Quantifying asymmetric parameter interaction. in sensitivity analysis: application to reservoir Quantifying asymmetric parameter interaction in sensitivity analysis: application to reservoir modeling Darryl Fenwick 1, Céline Scheidt 2 and Jef Caers 2 1 Streamsim Technologies, Inc, USA 2 Energy Resources

More information

Large Scale Modeling by Bayesian Updating Techniques

Large Scale Modeling by Bayesian Updating Techniques Large Scale Modeling by Bayesian Updating Techniques Weishan Ren Centre for Computational Geostatistics Department of Civil and Environmental Engineering University of Alberta Large scale models are useful

More information

Stochastic vs Deterministic Pre-stack Inversion Methods. Brian Russell

Stochastic vs Deterministic Pre-stack Inversion Methods. Brian Russell Stochastic vs Deterministic Pre-stack Inversion Methods Brian Russell Introduction Seismic reservoir analysis techniques utilize the fact that seismic amplitudes contain information about the geological

More information

Distance-based stochastic modeling: theory and applications

Distance-based stochastic modeling: theory and applications Distance-based stochastic modeling: theory and applications Jef Caers Stanford University, California, USA Energy Resources Engineering Department Abstract raditional to geostatistics is the quantification

More information

Entropy of Gaussian Random Functions and Consequences in Geostatistics

Entropy of Gaussian Random Functions and Consequences in Geostatistics Entropy of Gaussian Random Functions and Consequences in Geostatistics Paula Larrondo (larrondo@ualberta.ca) Department of Civil & Environmental Engineering University of Alberta Abstract Sequential Gaussian

More information

Modeling of Atmospheric Effects on InSAR Measurements With the Method of Stochastic Simulation

Modeling of Atmospheric Effects on InSAR Measurements With the Method of Stochastic Simulation Modeling of Atmospheric Effects on InSAR Measurements With the Method of Stochastic Simulation Z. W. LI, X. L. DING Department of Land Surveying and Geo-Informatics, Hong Kong Polytechnic University, Hung

More information

Advances in Locally Varying Anisotropy With MDS

Advances in Locally Varying Anisotropy With MDS Paper 102, CCG Annual Report 11, 2009 ( 2009) Advances in Locally Varying Anisotropy With MDS J.B. Boisvert and C. V. Deutsch Often, geology displays non-linear features such as veins, channels or folds/faults

More information

Geostatistical History Matching coupled with Adaptive Stochastic Sampling: A zonation-based approach using Direct Sequential Simulation

Geostatistical History Matching coupled with Adaptive Stochastic Sampling: A zonation-based approach using Direct Sequential Simulation Geostatistical History Matching coupled with Adaptive Stochastic Sampling: A zonation-based approach using Direct Sequential Simulation Eduardo Barrela* Instituto Superior Técnico, Av. Rovisco Pais 1,

More information

Formats for Expressing Acceptable Uncertainty

Formats for Expressing Acceptable Uncertainty Formats for Expressing Acceptable Uncertainty Brandon J. Wilde and Clayton V. Deutsch This short note aims to define a number of formats that could be used to express acceptable uncertainty. These formats

More information

Data Integration with Direct Multivariate Density Estimation

Data Integration with Direct Multivariate Density Estimation Data Integration with Direct Multivariate Density Estimation Sahyun Hong and Clayton V. Deutsch A longstanding problem in geostatistics is the integration of multiple secondary data in the construction

More information

Modeling Uncertainty in the Earth Sciences Jef Caers Stanford University

Modeling Uncertainty in the Earth Sciences Jef Caers Stanford University Probability theory and statistical analysis: a review Modeling Uncertainty in the Earth Sciences Jef Caers Stanford University Concepts assumed known Histograms, mean, median, spread, quantiles Probability,

More information

Bagging During Markov Chain Monte Carlo for Smoother Predictions

Bagging During Markov Chain Monte Carlo for Smoother Predictions Bagging During Markov Chain Monte Carlo for Smoother Predictions Herbert K. H. Lee University of California, Santa Cruz Abstract: Making good predictions from noisy data is a challenging problem. Methods

More information

Bayesian Methods for Machine Learning

Bayesian Methods for Machine Learning Bayesian Methods for Machine Learning CS 584: Big Data Analytics Material adapted from Radford Neal s tutorial (http://ftp.cs.utoronto.ca/pub/radford/bayes-tut.pdf), Zoubin Ghahramni (http://hunch.net/~coms-4771/zoubin_ghahramani_bayesian_learning.pdf),

More information

Net-to-gross from Seismic P and S Impedances: Estimation and Uncertainty Analysis using Bayesian Statistics

Net-to-gross from Seismic P and S Impedances: Estimation and Uncertainty Analysis using Bayesian Statistics Net-to-gross from Seismic P and S Impedances: Estimation and Uncertainty Analysis using Bayesian Statistics Summary Madhumita Sengupta*, Ran Bachrach, Niranjan Banik, esterngeco. Net-to-gross (N/G ) is

More information

The exact bootstrap method shown on the example of the mean and variance estimation

The exact bootstrap method shown on the example of the mean and variance estimation Comput Stat (2013) 28:1061 1077 DOI 10.1007/s00180-012-0350-0 ORIGINAL PAPER The exact bootstrap method shown on the example of the mean and variance estimation Joanna Kisielinska Received: 21 May 2011

More information

Econ 582 Nonparametric Regression

Econ 582 Nonparametric Regression Econ 582 Nonparametric Regression Eric Zivot May 28, 2013 Nonparametric Regression Sofarwehaveonlyconsideredlinearregressionmodels = x 0 β + [ x ]=0 [ x = x] =x 0 β = [ x = x] [ x = x] x = β The assume

More information

A033 PRACTICAL METHODS FOR UNCERTAINTY ASSESSMENT

A033 PRACTICAL METHODS FOR UNCERTAINTY ASSESSMENT A33 PRACTICAL METHODS FOR UNCERTAINTY ASSESSMENT OF FLOW PREDICTIONS FOR RESERVOIRS WITH SIGNIFICANT HISTORY AFIELD CASE STUDY ALEXANDRE CASTELLINl, JORGE L. LANDA, JITENDRA KIKANI 2 () ChevronTexaco,

More information

Latin Hypercube Sampling with Multidimensional Uniformity

Latin Hypercube Sampling with Multidimensional Uniformity Latin Hypercube Sampling with Multidimensional Uniformity Jared L. Deutsch and Clayton V. Deutsch Complex geostatistical models can only be realized a limited number of times due to large computational

More information

Appendix F. Computational Statistics Toolbox. The Computational Statistics Toolbox can be downloaded from:

Appendix F. Computational Statistics Toolbox. The Computational Statistics Toolbox can be downloaded from: Appendix F Computational Statistics Toolbox The Computational Statistics Toolbox can be downloaded from: http://www.infinityassociates.com http://lib.stat.cmu.edu. Please review the readme file for installation

More information

Acceptable Ergodic Fluctuations and Simulation of Skewed Distributions

Acceptable Ergodic Fluctuations and Simulation of Skewed Distributions Acceptable Ergodic Fluctuations and Simulation of Skewed Distributions Oy Leuangthong, Jason McLennan and Clayton V. Deutsch Centre for Computational Geostatistics Department of Civil & Environmental Engineering

More information

Application of Copulas as a New Geostatistical Tool

Application of Copulas as a New Geostatistical Tool Application of Copulas as a New Geostatistical Tool Presented by Jing Li Supervisors András Bardossy, Sjoerd Van der zee, Insa Neuweiler Universität Stuttgart Institut für Wasserbau Lehrstuhl für Hydrologie

More information

QUANTITATIVE INTERPRETATION

QUANTITATIVE INTERPRETATION QUANTITATIVE INTERPRETATION THE AIM OF QUANTITATIVE INTERPRETATION (QI) IS, THROUGH THE USE OF AMPLITUDE ANALYSIS, TO PREDICT LITHOLOGY AND FLUID CONTENT AWAY FROM THE WELL BORE This process should make

More information

Basics of Uncertainty Analysis

Basics of Uncertainty Analysis Basics of Uncertainty Analysis Chapter Six Basics of Uncertainty Analysis 6.1 Introduction As shown in Fig. 6.1, analysis models are used to predict the performances or behaviors of a product under design.

More information

Statistics Toolbox 6. Apply statistical algorithms and probability models

Statistics Toolbox 6. Apply statistical algorithms and probability models Statistics Toolbox 6 Apply statistical algorithms and probability models Statistics Toolbox provides engineers, scientists, researchers, financial analysts, and statisticians with a comprehensive set of

More information

Conditional Distribution Fitting of High Dimensional Stationary Data

Conditional Distribution Fitting of High Dimensional Stationary Data Conditional Distribution Fitting of High Dimensional Stationary Data Miguel Cuba and Oy Leuangthong The second order stationary assumption implies the spatial variability defined by the variogram is constant

More information

NONINFORMATIVE NONPARAMETRIC BAYESIAN ESTIMATION OF QUANTILES

NONINFORMATIVE NONPARAMETRIC BAYESIAN ESTIMATION OF QUANTILES NONINFORMATIVE NONPARAMETRIC BAYESIAN ESTIMATION OF QUANTILES Glen Meeden School of Statistics University of Minnesota Minneapolis, MN 55455 Appeared in Statistics & Probability Letters Volume 16 (1993)

More information

Automatic Determination of Uncertainty versus Data Density

Automatic Determination of Uncertainty versus Data Density Automatic Determination of Uncertainty versus Data Density Brandon Wilde and Clayton V. Deutsch It is useful to know how various measures of uncertainty respond to changes in data density. Calculating

More information

Density Estimation (II)

Density Estimation (II) Density Estimation (II) Yesterday Overview & Issues Histogram Kernel estimators Ideogram Today Further development of optimization Estimating variance and bias Adaptive kernels Multivariate kernel estimation

More information

Geostatistics for Seismic Data Integration in Earth Models

Geostatistics for Seismic Data Integration in Earth Models 2003 Distinguished Instructor Short Course Distinguished Instructor Series, No. 6 sponsored by the Society of Exploration Geophysicists European Association of Geoscientists & Engineers SUB Gottingen 7

More information

Prediction-Focused Subsurface Modeling: Investigating the Need for Accuracy in Flow-Based Inverse Modeling

Prediction-Focused Subsurface Modeling: Investigating the Need for Accuracy in Flow-Based Inverse Modeling Math Geosci (25) 47:73 9 DOI.7/s4-4-952-6 Prediction-Focused Subsurface Modeling: Investigating the Need for Accuracy in Flow-Based Inverse Modeling Céline Scheidt Philippe Renard Jef Caers Received: 2

More information

Nonparametric Bayesian Methods - Lecture I

Nonparametric Bayesian Methods - Lecture I Nonparametric Bayesian Methods - Lecture I Harry van Zanten Korteweg-de Vries Institute for Mathematics CRiSM Masterclass, April 4-6, 2016 Overview of the lectures I Intro to nonparametric Bayesian statistics

More information

A Program for Data Transformations and Kernel Density Estimation

A Program for Data Transformations and Kernel Density Estimation A Program for Data Transformations and Kernel Density Estimation John G. Manchuk and Clayton V. Deutsch Modeling applications in geostatistics often involve multiple variables that are not multivariate

More information

Conditional Standardization: A Multivariate Transformation for the Removal of Non-linear and Heteroscedastic Features

Conditional Standardization: A Multivariate Transformation for the Removal of Non-linear and Heteroscedastic Features Conditional Standardization: A Multivariate Transformation for the Removal of Non-linear and Heteroscedastic Features Ryan M. Barnett Reproduction of complex multivariate features, such as non-linearity,

More information

Exercises Tutorial at ICASSP 2016 Learning Nonlinear Dynamical Models Using Particle Filters

Exercises Tutorial at ICASSP 2016 Learning Nonlinear Dynamical Models Using Particle Filters Exercises Tutorial at ICASSP 216 Learning Nonlinear Dynamical Models Using Particle Filters Andreas Svensson, Johan Dahlin and Thomas B. Schön March 18, 216 Good luck! 1 [Bootstrap particle filter for

More information

Chapter 12 PAWL-Forced Simulated Tempering

Chapter 12 PAWL-Forced Simulated Tempering Chapter 12 PAWL-Forced Simulated Tempering Luke Bornn Abstract In this short note, we show how the parallel adaptive Wang Landau (PAWL) algorithm of Bornn et al. (J Comput Graph Stat, to appear) can be

More information

STAT 518 Intro Student Presentation

STAT 518 Intro Student Presentation STAT 518 Intro Student Presentation Wen Wei Loh April 11, 2013 Title of paper Radford M. Neal [1999] Bayesian Statistics, 6: 475-501, 1999 What the paper is about Regression and Classification Flexible

More information

Global Sensitivity Analysis of Complex Systems implications for natural resources. Jef Caers Geological Sciences Stanford University, USA

Global Sensitivity Analysis of Complex Systems implications for natural resources. Jef Caers Geological Sciences Stanford University, USA Global Sensitivity Analysis of Complex Systems implications for natural resources Jef Caers Geological Sciences Stanford University, USA Analysis of (Subsurface) Systems The model: description of the world

More information

MACHINE LEARNING FOR PRODUCTION FORECASTING: ACCURACY THROUGH UNCERTAINTY

MACHINE LEARNING FOR PRODUCTION FORECASTING: ACCURACY THROUGH UNCERTAINTY MACHINE LEARNING FOR PRODUCTION FORECASTING: ACCURACY THROUGH UNCERTAINTY 7 TH RESERVES ESTIMATION UNCONVENTIONALS JUNE 20 22, 2017 HOUSTON, TX DAVID FULFORD APACHE CORPORATION PRODUCTION FORECASTING IN

More information

The Jackknife-Like Method for Assessing Uncertainty of Point Estimates for Bayesian Estimation in a Finite Gaussian Mixture Model

The Jackknife-Like Method for Assessing Uncertainty of Point Estimates for Bayesian Estimation in a Finite Gaussian Mixture Model Thai Journal of Mathematics : 45 58 Special Issue: Annual Meeting in Mathematics 207 http://thaijmath.in.cmu.ac.th ISSN 686-0209 The Jackknife-Like Method for Assessing Uncertainty of Point Estimates for

More information

Bootstrap, Jackknife and other resampling methods

Bootstrap, Jackknife and other resampling methods Bootstrap, Jackknife and other resampling methods Part III: Parametric Bootstrap Rozenn Dahyot Room 128, Department of Statistics Trinity College Dublin, Ireland dahyot@mee.tcd.ie 2005 R. Dahyot (TCD)

More information

Quantifying uncertainty of geological 3D layer models, constructed with a-priori

Quantifying uncertainty of geological 3D layer models, constructed with a-priori Quantifying uncertainty of geological 3D layer models, constructed with a-priori geological expertise Jan Gunnink, Denise Maljers 2 and Jan Hummelman 2, TNO Built Environment and Geosciences Geological

More information

Anomaly Density Estimation from Strip Transect Data: Pueblo of Isleta Example

Anomaly Density Estimation from Strip Transect Data: Pueblo of Isleta Example Anomaly Density Estimation from Strip Transect Data: Pueblo of Isleta Example Sean A. McKenna, Sandia National Laboratories Brent Pulsipher, Pacific Northwest National Laboratory May 5 Distribution Statement

More information

Computational Challenges in Reservoir Modeling. Sanjay Srinivasan The Pennsylvania State University

Computational Challenges in Reservoir Modeling. Sanjay Srinivasan The Pennsylvania State University Computational Challenges in Reservoir Modeling Sanjay Srinivasan The Pennsylvania State University Well Data 3D view of well paths Inspired by an offshore development 4 platforms 2 vertical wells 2 deviated

More information

Experimental Design and Data Analysis for Biologists

Experimental Design and Data Analysis for Biologists Experimental Design and Data Analysis for Biologists Gerry P. Quinn Monash University Michael J. Keough University of Melbourne CAMBRIDGE UNIVERSITY PRESS Contents Preface page xv I I Introduction 1 1.1

More information

Oil Field Production using Machine Learning. CS 229 Project Report

Oil Field Production using Machine Learning. CS 229 Project Report Oil Field Production using Machine Learning CS 229 Project Report Sumeet Trehan, Energy Resources Engineering, Stanford University 1 Introduction Effective management of reservoirs motivates oil and gas

More information

Facies Modeling in Presence of High Resolution Surface-based Reservoir Models

Facies Modeling in Presence of High Resolution Surface-based Reservoir Models Facies Modeling in Presence of High Resolution Surface-based Reservoir Models Kevin Zhang Centre for Computational Geostatistics Department of Civil and Environmental Engineering University of Alberta

More information

Quantitative Seismic Interpretation An Earth Modeling Perspective

Quantitative Seismic Interpretation An Earth Modeling Perspective Quantitative Seismic Interpretation An Earth Modeling Perspective Damien Thenin*, RPS, Calgary, AB, Canada TheninD@rpsgroup.com Ron Larson, RPS, Calgary, AB, Canada LarsonR@rpsgroup.com Summary Earth models

More information

USING GEOSTATISTICS TO DESCRIBE COMPLEX A PRIORI INFORMATION FOR INVERSE PROBLEMS THOMAS M. HANSEN 1,2, KLAUS MOSEGAARD 2 and KNUD S.

USING GEOSTATISTICS TO DESCRIBE COMPLEX A PRIORI INFORMATION FOR INVERSE PROBLEMS THOMAS M. HANSEN 1,2, KLAUS MOSEGAARD 2 and KNUD S. USING GEOSTATISTICS TO DESCRIBE COMPLEX A PRIORI INFORMATION FOR INVERSE PROBLEMS THOMAS M. HANSEN 1,2, KLAUS MOSEGAARD 2 and KNUD S. CORDUA 1 1 Institute of Geography & Geology, University of Copenhagen,

More information

4th HR-HU and 15th HU geomathematical congress Geomathematics as Geoscience Reliability enhancement of groundwater estimations

4th HR-HU and 15th HU geomathematical congress Geomathematics as Geoscience Reliability enhancement of groundwater estimations Reliability enhancement of groundwater estimations Zoltán Zsolt Fehér 1,2, János Rakonczai 1, 1 Institute of Geoscience, University of Szeged, H-6722 Szeged, Hungary, 2 e-mail: zzfeher@geo.u-szeged.hu

More information

Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful?

Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful? Journal of Modern Applied Statistical Methods Volume 10 Issue Article 13 11-1-011 Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful?

More information

Multitask Learning of Environmental Spatial Data

Multitask Learning of Environmental Spatial Data 9th International Congress on Environmental Modelling and Software Brigham Young University BYU ScholarsArchive 6th International Congress on Environmental Modelling and Software - Leipzig, Germany - July

More information

A NEW APPROACH FOR QUANTIFYING THE IMPACT OF GEOSTATISTICAL UNCERTAINTY ON PRODUCTION FORECASTS: THE JOINT MODELING METHOD

A NEW APPROACH FOR QUANTIFYING THE IMPACT OF GEOSTATISTICAL UNCERTAINTY ON PRODUCTION FORECASTS: THE JOINT MODELING METHOD A NEW APPROACH FOR QUANTIFYING THE IMPACT OF GEOSTATISTICAL UNCERTAINTY ON PRODUCTION FORECASTS: THE JOINT MODELING METHOD IAMG, Cancun, September 6-1, 001 Isabelle Zabalza-Mezghani, IFP Emmanuel Manceau,

More information

Estimation of Parameters

Estimation of Parameters CHAPTER Probability, Statistics, and Reliability for Engineers and Scientists FUNDAMENTALS OF STATISTICAL ANALYSIS Second Edition A. J. Clark School of Engineering Department of Civil and Environmental

More information

Statistics - Lecture One. Outline. Charlotte Wickham 1. Basic ideas about estimation

Statistics - Lecture One. Outline. Charlotte Wickham  1. Basic ideas about estimation Statistics - Lecture One Charlotte Wickham wickham@stat.berkeley.edu http://www.stat.berkeley.edu/~wickham/ Outline 1. Basic ideas about estimation 2. Method of Moments 3. Maximum Likelihood 4. Confidence

More information

Estimation of Operational Risk Capital Charge under Parameter Uncertainty

Estimation of Operational Risk Capital Charge under Parameter Uncertainty Estimation of Operational Risk Capital Charge under Parameter Uncertainty Pavel V. Shevchenko Principal Research Scientist, CSIRO Mathematical and Information Sciences, Sydney, Locked Bag 17, North Ryde,

More information

3D geostatistical porosity modelling: A case study at the Saint-Flavien CO 2 storage project

3D geostatistical porosity modelling: A case study at the Saint-Flavien CO 2 storage project 3D geostatistical porosity modelling: A case study at the Saint-Flavien CO 2 storage project Maxime Claprood Institut national de la recherche scientifique, Québec, Canada Earth Modelling 2013 October

More information

Linear inverse Gaussian theory and geostatistics a tomography example København Ø,

Linear inverse Gaussian theory and geostatistics a tomography example København Ø, Linear inverse Gaussian theory and geostatistics a tomography example Thomas Mejer Hansen 1, ndre Journel 2, lbert Tarantola 3 and Klaus Mosegaard 1 1 Niels Bohr Institute, University of Copenhagen, Juliane

More information

On Modifications to Linking Variance Estimators in the Fay-Herriot Model that Induce Robustness

On Modifications to Linking Variance Estimators in the Fay-Herriot Model that Induce Robustness Statistics and Applications {ISSN 2452-7395 (online)} Volume 16 No. 1, 2018 (New Series), pp 289-303 On Modifications to Linking Variance Estimators in the Fay-Herriot Model that Induce Robustness Snigdhansu

More information

The Necessity of a Multiple-Point Prior Model

The Necessity of a Multiple-Point Prior Model The Necessity of a Multiple-Point Prior Model Andre Journel (1) and Tuanfeng Zhang (1) (1) Department of Geological and Environmental Sciences, Stanford University, CA 94305, U.S.A. email: journel@pangea.stanford.edu

More information

Reservoir Modeling with GSLIB. Overview

Reservoir Modeling with GSLIB. Overview Reservoir Modeling with GSLIB Overview Objectives of the Course What is Geostatistics? Why Geostatistics / 3-D Modeling? Uncertainty Quantification and Decision Making Heterogeneous Reservoir Modeling

More information

Bootstrapping the Confidence Intervals of R 2 MAD for Samples from Contaminated Standard Logistic Distribution

Bootstrapping the Confidence Intervals of R 2 MAD for Samples from Contaminated Standard Logistic Distribution Pertanika J. Sci. & Technol. 18 (1): 209 221 (2010) ISSN: 0128-7680 Universiti Putra Malaysia Press Bootstrapping the Confidence Intervals of R 2 MAD for Samples from Contaminated Standard Logistic Distribution

More information

IJMGE Int. J. Min. & Geo-Eng. Vol.49, No.1, June 2015, pp

IJMGE Int. J. Min. & Geo-Eng. Vol.49, No.1, June 2015, pp IJMGE Int. J. Min. & Geo-Eng. Vol.49, No.1, June 2015, pp.131-142 Joint Bayesian Stochastic Inversion of Well Logs and Seismic Data for Volumetric Uncertainty Analysis Moslem Moradi 1, Omid Asghari 1,

More information

Porosity prediction using cokriging with multiple secondary datasets

Porosity prediction using cokriging with multiple secondary datasets Cokriging with Multiple Attributes Porosity prediction using cokriging with multiple secondary datasets Hong Xu, Jian Sun, Brian Russell, Kris Innanen ABSTRACT The prediction of porosity is essential for

More information

Computer Science, Informatik 4 Communication and Distributed Systems. Simulation. Discrete-Event System Simulation. Dr.

Computer Science, Informatik 4 Communication and Distributed Systems. Simulation. Discrete-Event System Simulation. Dr. Simulation Discrete-Event System Simulation Chapter 0 Output Analysis for a Single Model Purpose Objective: Estimate system performance via simulation If θ is the system performance, the precision of the

More information

On the Optimal Scaling of the Modified Metropolis-Hastings algorithm

On the Optimal Scaling of the Modified Metropolis-Hastings algorithm On the Optimal Scaling of the Modified Metropolis-Hastings algorithm K. M. Zuev & J. L. Beck Division of Engineering and Applied Science California Institute of Technology, MC 4-44, Pasadena, CA 925, USA

More information

Algorithm-Independent Learning Issues

Algorithm-Independent Learning Issues Algorithm-Independent Learning Issues Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr CS 551, Spring 2007 c 2007, Selim Aksoy Introduction We have seen many learning

More information

Chapter 11. Output Analysis for a Single Model Prof. Dr. Mesut Güneş Ch. 11 Output Analysis for a Single Model

Chapter 11. Output Analysis for a Single Model Prof. Dr. Mesut Güneş Ch. 11 Output Analysis for a Single Model Chapter Output Analysis for a Single Model. Contents Types of Simulation Stochastic Nature of Output Data Measures of Performance Output Analysis for Terminating Simulations Output Analysis for Steady-state

More information

7 Geostatistics. Figure 7.1 Focus of geostatistics

7 Geostatistics. Figure 7.1 Focus of geostatistics 7 Geostatistics 7.1 Introduction Geostatistics is the part of statistics that is concerned with geo-referenced data, i.e. data that are linked to spatial coordinates. To describe the spatial variation

More information

Machine Learning Applied to 3-D Reservoir Simulation

Machine Learning Applied to 3-D Reservoir Simulation Machine Learning Applied to 3-D Reservoir Simulation Marco A. Cardoso 1 Introduction The optimization of subsurface flow processes is important for many applications including oil field operations and

More information

Abstract. Introduction. Cheolkyun Jeong, Céline Scheidt, Jef Caers, and Tapan Mukerji Department of Energy Resources Engineering Stanford University

Abstract. Introduction. Cheolkyun Jeong, Céline Scheidt, Jef Caers, and Tapan Mukerji Department of Energy Resources Engineering Stanford University 1 Modeling Geological Scenario Uncertainty from Seismic Data using Pattern Similarity: Application to a Channelized Turbidite Reservoir in Equatorial Guinea, West Africa Cheolkyun Jeong, Céline Scheidt,

More information

Multiple realizations using standard inversion techniques a

Multiple realizations using standard inversion techniques a Multiple realizations using standard inversion techniques a a Published in SEP report, 105, 67-78, (2000) Robert G Clapp 1 INTRODUCTION When solving a missing data problem, geophysicists and geostatisticians

More information

Development of Stochastic Artificial Neural Networks for Hydrological Prediction

Development of Stochastic Artificial Neural Networks for Hydrological Prediction Development of Stochastic Artificial Neural Networks for Hydrological Prediction G. B. Kingston, M. F. Lambert and H. R. Maier Centre for Applied Modelling in Water Engineering, School of Civil and Environmental

More information

Stepwise Conditional Transformation for Simulation of Multiple Variables 1

Stepwise Conditional Transformation for Simulation of Multiple Variables 1 Mathematical Geology, Vol. 35, No. 2, February 2003 ( C 2003) Stepwise Conditional Transformation for Simulation of Multiple Variables 1 Oy Leuangthong 2 and Clayton V. Deutsch 2 Most geostatistical studies

More information

Thomas Bayes versus the wedge model: An example inference using a geostatistical prior function

Thomas Bayes versus the wedge model: An example inference using a geostatistical prior function Thomas Bayes versus the wedge model: An example inference using a geostatistical prior function Jason M. McCrank, Gary F. Margrave, and Don C. Lawton ABSTRACT The Bayesian inference is used to estimate

More information

Building Blocks for Direct Sequential Simulation on Unstructured Grids

Building Blocks for Direct Sequential Simulation on Unstructured Grids Building Blocks for Direct Sequential Simulation on Unstructured Grids Abstract M. J. Pyrcz (mpyrcz@ualberta.ca) and C. V. Deutsch (cdeutsch@ualberta.ca) University of Alberta, Edmonton, Alberta, CANADA

More information

Adaptive Dual Control

Adaptive Dual Control Adaptive Dual Control Björn Wittenmark Department of Automatic Control, Lund Institute of Technology Box 118, S-221 00 Lund, Sweden email: bjorn@control.lth.se Keywords: Dual control, stochastic control,

More information

LINKING GEOSTATISTICS WITH BASIN AND PETROLEUM SYSTEM MODELING: ASSESSMENT OF SPATIAL UNCERTAINTIES

LINKING GEOSTATISTICS WITH BASIN AND PETROLEUM SYSTEM MODELING: ASSESSMENT OF SPATIAL UNCERTAINTIES LINKING GEOSTATISTICS WITH BASIN AND PETROLEUM SYSTEM MODELING: ASSESSMENT OF SPATIAL UNCERTAINTIES A REPORT SUBMITTED TO THE DEPARTMENT OF ENERGY RESOURCES ENGINEERING OF STANFORD UNIVERSITY IN PARTIAL

More information

Confidence Intervals in Ridge Regression using Jackknife and Bootstrap Methods

Confidence Intervals in Ridge Regression using Jackknife and Bootstrap Methods Chapter 4 Confidence Intervals in Ridge Regression using Jackknife and Bootstrap Methods 4.1 Introduction It is now explicable that ridge regression estimator (here we take ordinary ridge estimator (ORE)

More information

Assessment of Hydraulic Conductivity Upscaling Techniques and. Associated Uncertainty

Assessment of Hydraulic Conductivity Upscaling Techniques and. Associated Uncertainty CMWRXVI Assessment of Hydraulic Conductivity Upscaling Techniques and Associated Uncertainty FARAG BOTROS,, 4, AHMED HASSAN 3, 4, AND GREG POHLL Division of Hydrologic Sciences, University of Nevada, Reno

More information

Introduction to machine learning and pattern recognition Lecture 2 Coryn Bailer-Jones

Introduction to machine learning and pattern recognition Lecture 2 Coryn Bailer-Jones Introduction to machine learning and pattern recognition Lecture 2 Coryn Bailer-Jones http://www.mpia.de/homes/calj/mlpr_mpia2008.html 1 1 Last week... supervised and unsupervised methods need adaptive

More information

Chapter 2: Resampling Maarten Jansen

Chapter 2: Resampling Maarten Jansen Chapter 2: Resampling Maarten Jansen Randomization tests Randomized experiment random assignment of sample subjects to groups Example: medical experiment with control group n 1 subjects for true medicine,

More information