a coherent structure approach for parameter estimation in
play

A coherent structure approach for parameter estimation in Data - PowerPoint PPT Presentation

A coherent structure approach for parameter estimation in Data Assimilation John Maclean 1 , Naratip Santitissadeekorn 2 , Christopher KRT Jones 1 1 Department of Mathematics and RENCI, University of North Carolina at Chapel Hill 2 Department of


  1. A coherent structure approach for parameter estimation in Data Assimilation John Maclean 1 , Naratip Santitissadeekorn 2 , Christopher KRT Jones 1 1 Department of Mathematics and RENCI, University of North Carolina at Chapel Hill 2 Department of Mathematics, University of Surrey, Guildford June 13, 2017 J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 1 / 17

  2. Data Assimilation setup We consider x, θ to have come from a model ˙ θ = 0 , x = f ( x, θ, t ) ˙ with the interpretation that θ represents parameters and x represents tracers in the flow . Our goal is to estimate the parameters θ affecting a flow, given observations x o of passive tracers in the flow. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 2 / 17

  3. Data Assimilation setup We consider x, θ to have come from a model ˙ θ = 0 , x = f ( x, θ, t ) ˙ with the interpretation that θ represents parameters and x represents tracers in the flow . Our goal is to estimate the parameters θ affecting a flow, given observations x o of passive tracers in the flow. As usual we employ Bayes’ rule to formulate the data assimilation problem: p ( θ, x | x o ) ∝ p ( x o | θ, x ) p ( θ, x ) J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 2 / 17

  4. (Standard, naive) particle filters The particle filter sequentially approximates the distribution of θ at time n by a set { θ ( i ) n , w ( i ) n } , i = 1 , . . . , N of particles and weights. The weight update for the i -th particle is n ∝ w ( i ) w ( i ) n | x ( i ) n , θ ( i ) n − 1 p ( x o n ) � � − 1 ≈ w ( i ) n ) T R − 1 ( x o n − x ( i ) n − x ( i ) 2( x o n − 1 exp n ) , where R is the covariance matrix for the observations. The key quantity by which the particle filter gains information is the innovation, n − x ( i ) x o n . J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 3 / 17

  5. Toy model We consider a kinematic traveling wave model in the co-moving frame, perturbed by an oscillatory disturbance and stochastically perturbed in the x 1 -direction: dx 1 = c − A sin( Kx 1 ) cos( x 2 ) + εl 1 sin( k 1 ( x 1 − c 1 t )) cos( l 1 x 2 ) + σdW (1) dx 2 = AK cos( Kx 1 ) sin( x 2 ) + εk 1 cos( k 1 ( x 1 − c 1 t )) sin( l 1 x 2 ) . (2) We will perform experiments to attempt to ‘discover’ the true values of ε , and/or k 1 , given all the other parameters are fixed and given observations from a run with the ‘true’ parameters. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 4 / 17

  6. Toy model We consider a kinematic traveling wave model in the co-moving frame, perturbed by an oscillatory disturbance and stochastically perturbed in the x 1 -direction: dx 1 = c − A sin( Kx 1 ) cos( x 2 ) + εl 1 sin( k 1 ( x 1 − c 1 t )) cos( l 1 x 2 ) + σdW (1) dx 2 = AK cos( Kx 1 ) sin( x 2 ) + εk 1 cos( k 1 ( x 1 − c 1 t )) sin( l 1 x 2 ) . (2) We will perform experiments to attempt to ‘discover’ the true values of ε , and/or k 1 , given all the other parameters are fixed and given observations from a run with the ‘true’ parameters. This flow contains two gyres. For the value ε = 0 . 3 that we choose to be the truth, tracer trajectories inside the gyres are dominated by chaotic advection on long time scales. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 4 / 17

  7. Particle Filter innovation What will the particle filter do if we initialise all tracers in the boundaries of the gyres? Let us look at the innovations... 1.5 1.6 1.4 n ) /N n ) /N 1.2 n − x ( i ) 1 n − x ( i ) 1 i =0 ( x o i =0 ( x o 0.8 0.5 � N � N 0.6 0.4 0 0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Particle value Particle value (a) Observation taken at t = 20 . (b) Observation taken at t = 30 . The figures show experiments in which we uniformly spaced 2000 guesses for ε in [0 , 1] , numerically integrated 50 tracers using each value of ε , and calculated the innovation. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 5 / 17

  8. Coherent patterns We would like to perform Data Assimilation by assimilating the pattern of the observed tracers. In so doing, we hope to exploit the robustness of coherent patterns to peturbations, while preserving whatever information the tracers carry on the model parameters. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 6 / 17

  9. Coherent patterns Almost-invariant sets Trajectories stay in the almost-invariant sets for a comparatively long time before escaping to another region [Dellnitz and Jung 99]. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 7 / 17

  10. Coherent patterns Coherent sets Coherent sets are regions in state space, for example a coherent vortex or nonlinear jet, that move along with the flow without dispersing. Coherent sets must also be robust under small diffusive peturbations. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 8 / 17

  11. Coherent patterns Usage in Data Assimilation We rely on numerical methods that extract the coherent pattern from data, so that we can produce both a ‘simulated’ pattern (extracted from simulated tracer positions) and an ‘observed’ pattern (extracted from data). We use PCA to find the almost-invariant sets, shown here for many tracers: and for few tracers: Intial position at t=0 At t=20 (Realization 1) 23 3 8 3 1 2 7 3 21 14 20 9 2.5 2.5 10 12 19 8 16 17 17 15 11 23 12 16 2 2 22 20 5 25 18 7 3 25 1 y y 1.5 6 14 1.5 9 4 2 19 15 21 1 1 18 13 24 24 10 0.5 0.5 6 22 11 5 13 4 0 0 0 1 2 3 4 5 6 0 1 2 3 4 5 6 x x J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 9 / 17

  12. Assimilating patterns, p1 Suppose the random variable y o represents the observed coherent pattern. The updated statement of Bayes’ rule is that p ( θ | y o ) ∝ p ( y o | θ ) p ( θ ) , where � p ( y o | θ ) = p ( y o | θ, x 0: n ) p ( x 0: n | θ ) dx 0: n ...is hard to evaluate. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 10 / 17

  13. Assimilating patterns, p1 Suppose the random variable y o represents the observed coherent pattern. The updated statement of Bayes’ rule is that p ( θ | y o ) ∝ p ( y o | θ ) p ( θ ) , where � p ( y o | θ ) = p ( y o | θ, x 0: n ) p ( x 0: n | θ ) dx 0: n ...is hard to evaluate. We proceed by replacing x 0: n in the likelihood function in the integral with ˆ x 0: n ( θ ) , a realisation of x 0: n given θ , to obtain p ( y o | θ ) ≈ p ( y o | θ, ˆ x 0: n ( θ )) . J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 10 / 17

  14. Assimilating patterns, p2 Unfortunately we still cannot calculate p ( y | θ, ˆ x 0: n ( θ )) . We turn instead to an Approximate Bayesian Computation (Rubin, 1984; Sisson, Fan, Tanaka, 2007), which uses a distance function ρ to substitute for the likelihood function. A basic algorithm description is Algorithm 0 Step 1. Sample θ ∼ p ( θ ) Step 2. Sample y ∼ p ( y | θ ) Step 3. Accept θ if ρ ( y, y o ) ≤ ε. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 11 / 17

  15. SMC-ABC Algorithm 0 Step 1. Sample θ ∼ p ( θ ) Step 2. Sample y ∼ p ( y | θ ) Step 3. Accept θ if ρ ( y, y o ) ≤ ε. In this way all ABC algorithms sample from p ε ( θ, y | y o ) ∝ p ( y | θ ) p ( θ ) I ε ( y ) , where � for ρ ( y, y o ) < ε, 1 I ε ( y ) = 0 otherwise . We choose to use the Hellinger distance for ρ , so the fundamental source of information in our ABC scheme is the Hellinger distance between an observed pattern and a simulated pattern. J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 12 / 17

  16. Building blocks for DA schemes - innovation vs Hellinger distance 0.35 0.28 0.26 0.3 Hellinger distance Hellinger distance 0.24 0.22 0.25 0.2 0.2 0.18 0.16 0.15 0.14 0.12 0.1 0.1 0.05 0.08 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Particle value Particle value 1.5 1.6 1.4 n ) /N n ) /N 1.2 n − x ( i ) 1 n − x ( i ) 1 i =0 ( x o i =0 ( x o 0.8 0.5 � N � N 0.6 0.4 0 0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Particle value Particle value (a) t=20 (b) t=30 We repeat the prior experiment, now including results for the Hellinger distance between the observed pattern and the simulated pattern at each value of ε . J. Maclean, N. Santitissadeekorn, C.K.R.T. Jones Coherent structures and LaDA June 13, 2017 13 / 17

Recommend


More recommend