Example: Ground Motion Attenuation

Transcription

1 Example: Ground Motion Attenuation Problem: Predict the probability distribution for Peak Ground Acceleration (PGA), the level of ground shaking caused by an earthquake Earthquake records are used to update the predictive probability for PGA based on earthquake magnitude M, distance to the fault d, and local soil conditions G= ( G, G ) (soil classified A, B or C) Well-known attenuation relation developed by Boore, Joyner and 2 2 Fumal (BJF, 1993). Let R= d + h, h= fictitious depth: 2 ( ) ( ) ( ) ( ) log PGA = b + b M 6 + b M 6 + b R + b log R + b G + b G + ε B 7 qmdgbh (,, ;, ) + ε (where b= ( b1,..., b7)) ε = uncertain prediction error modelled as (0, σ) (max. entropy model) B C C

2 Example: Ground Motion Attenuation Goal: Estimate p( PGA U, D, M) for input U = ( M, d, G) D Available set of data where U = ( M, d, G ) n n n n (magnitude, distance, soil conditions), and corresponding Yn = ( PGA) n are data from earthquakes at various sites (we use =271 ground motion records from 20 earthquakes) Model class M is from BJF model with specified prior PDF over the model parameters θ = (,, bhσ) Robust posterior predictive probability model: (,D, ) = (, ) (, ) M D M = { U, Y : n= 1,..., } ppgauθ (, ) n p PGA U p PGA U θ p θ dθ n

3 Bayesian Updating Bayes Theorem: p θ D, = cp θ M M p Yn Un, θ n= 1 Computing Optimal Posterior Predictive Model ( ) ( ) ( ) Find optimal (most probable values) of parameters which 1 maximize, then to (Laplace s asymptotic approx.): ( D M) p θ, (, DM, ) ppgauθ (, ˆ) ppgau O( ) Assumes M is globally identifiable on ( is unique) and need a large amount of data for an acceptable approximation (updated PDF will then have single sharp peak). Then no need to evaluate and is insensitive to the choice of prior PDF θˆ Parameter estimation (i.e. using ) is reasonable only under these conditions; otherwise, spurious reduction in uncertainty in predictions θˆ D ˆ θ θˆ c

4 Bayesian Updating (Continued) Computing Robust Posterior Predictive Model c ormalizing constant, and p θ,, is difficult to evaluate However, if we can generate M samples from p( θ D ), M, then we can approximate the robust PDF by the corresponding sample mean: (, D, ) = (, ) (, ) M D M p PGA U p PGA U θ p θ dθ 1 M M k= 1 ( ) D M { θ, k = 1,, M } k (, θ ) k ppgau These samples can be obtained using stochastic simulation methods, e.g. Gibbs sampler, Metropolis-Hastings (more later)

5 Sampling Posterior (Updated) PDF Samples generated using Markov Chain Monte Carlo Prior PDF chosen to reflect knowledge (or lack thereof) For the regression coefficients b, take i.i.d. (0,10), i.e. each has zero mean and standard deviation of 10 (very flat) For the depth parameter h (km), lognormal distribution with mean and variance based on the depth of earthquakes in the data set b i s h

6 MCMC Samples from Posterior PDF Model Class 4 b 1 b 2 b 5 b 6 b 7 h

7 Posterior Samples: Model Class 4 b 2 b 5 b 6 b 7 h b 1 b 2 b 5 x BJF 93 < 1 σ < 2 σ > 2 σ b 6 b 7

8 Optimal vs. Robust Predictive Analysis Results using MCMC samples for robust PDF compared to results for optimal PDF computed from values reported in BJF 93 User input U=(M, d, G) Magnitude 7.0 Distance to fault 30km Site geology is stiff soil

9 Posterior Predictive Analysis 1971 San Fernando M 6.6 JPL Building km 1987 Whittier arrows M Old House Rd 19.0 km 1991 Sierra Madre M S Wilson Ave 18.1 km 1994 orthridge M Sierra Madre Villa 36.2 km

10 Model Class Selection Bayesian model class selection (Beck and Yuen 2004) for set M of candidate models p( D Mi) P( Mi M) P( M D,M) =, i= 1,..., I i I j= 1 p ( D M j) P( M j M) The evidence for model class M i is given by p ( D M ) p( D M, θ ) p( θ M ) dθ = i i i i i i Evaluating the evidence directly by stochastic simulation would require sampling from the prior, which is typically inefficient

11 Model Class Selection (Continued) The evidence may expressed as ( D Mi ) p( D M θ ) p( θ M ) p( θ M D) dθ H p( θ M D) ln p = ln i, i i i i i, i + i i, First term may be approximated from MCMC samples k = 1 ( D M ) ( M ) ( M D) ln p i, θi p θi i p θi i, dθi 1 ln p( D M, ˆ ) ( ˆ ) i θk p θ k Mi umerous methods for approximating information entropy from samples

12 Model Class Selection (Continued) ew result generalizes to any model class the conclusion by Beck and Yuen (2004) made for a globally identifiable model class using an asymptotic approximation of the evidence for the model class: ( D Mi ) ( D M ) ( M ) ( M D) p( θ M D) p( θ M D) dθ ln p = ln p, θ p θ p θ, dθ ln,, i i i i i i i i i i i i ( θ i Mi, D) p ( θ M ) p = ln p ( D Mi, θ i) p( θi Mi, D) dθi ln p( θi Mi, D) dθi i i = Data Fit - Expected Information Gained from Data

13 Model Class Selection ( ) ( ) ( ) ( ) PGA = b1 + b2 M + b3 M + b4 R + b5 10 R + b6gb + b7gc + ε R = d + h log 6 6 log, Model Class b 1 b 2 b 3 b 4 b 5 b 6 b 7 h σ Prob. (%) BJF Model (0.327) (0.034) (0.039) (0.002) (0.252) (0.038) (2.631) Model (0.326) (0.023) (0.002) (0.254) (0.037) (2.534) Model (0.103) (0.034) (0.040) (0.064) (0.038) (1.510) Model (0.108) (0.023) (0.063) (1.162) Model (0.043) (0.026) (0.001) (0.039) (0.041) (1.664) Model (0.114) (0.025) (0.072) (1.720) Model (0.022) (0.021) (0.033) (0.034) (1.041)

14 Comparison by Mean Data Fit Only ( ) ( ) ( ) ( ) PGA = b1 + b2 M + b3 M + b4 R + b5 10 R + b6gb + b7gc + ε R = d + h log 6 6 log, Model Class b 1 b 2 b 3 b 4 b 5 b 6 b 7 h σ Data Fit (%) BJF Model (0.327) (0.034) (0.039) (0.002) (0.252) (0.038) (2.631) Model (0.326) (0.023) (0.002) (0.254) (0.037) (2.534) Model (0.103) (0.034) (0.040) (0.064) (0.038) (1.510) Model (0.108) (0.023) (0.063) (1.162) Model (0.043) (0.026) (0.001) (0.039) (0.041) (1.664) Model (0.114) (0.025) (0.072) (1.720) Model (0.022) (0.021) (0.033) (0.034) (1.041)

15 Posterior Samples : Model Class 7 b 5 b 6 b 7 h b 2 b 5 x BJF 93 < 1 σ < 2 σ > 2 σ b 6 b 7