Geolocation of Icelandic Cod using a modified Particle Filter Method David Brickman Vilhjamur Thorsteinsson.

Slides:



Advertisements
Similar presentations
Bayesian Belief Propagation
Advertisements

Example Project and Numerical Integration Computational Neuroscience 03 Lecture 11.
Lindsey Bleimes Charlie Garrod Adam Meyerson
Jose-Luis Blanco, Javier González, Juan-Antonio Fernández-Madrigal University of Málaga (Spain) Dpt. of System Engineering and Automation May Pasadena,
EKF, UKF TexPoint fonts used in EMF.
Simulated Annealing General Idea: Start with an initial solution
CS479/679 Pattern Recognition Dr. George Bebis
The Capital Budgeting Decision (Chapter 12)  Capital Budgeting: An Overview  Estimating Incremental Cash Flows  Payback Period  Net Present Value 
1 12. Principles of Parameter Estimation The purpose of this lecture is to illustrate the usefulness of the various concepts introduced and studied in.
Acoustic design by simulated annealing algorithm
Budapest May 27, 2008 Unifying mixed linear models and the MASH algorithm for breakpoint detection and correction Anders Grimvall, Sackmone Sirisack, Agne.
Defining alternative climatologies: why, what for and how.
Your Name Your Title Your Organization (Line #1) Your Organization (Line #2) Semester 2 Update Joe Hoatam Josh Merritt Aaron Nielsen.
CHAPTER 8 A NNEALING- T YPE A LGORITHMS Organization of chapter in ISSO –Introduction to simulated annealing –Simulated annealing algorithm Basic algorithm.
1 アンサンブルカルマンフィルターによ る大気海洋結合モデルへのデータ同化 On-line estimation of observation error covariance for ensemble-based filters Genta Ueno The Institute of Statistical.
Aspects of Conditional Simulation and estimation of hydraulic conductivity in coastal aquifers" Luit Jan Slooten.
Planning under Uncertainty
“Estimates of (steric) SSH rise from ocean syntheses" Detlef Stammer Universität Hamburg  SODA (J. Carton)  AWI roWE (J. Schroeter, M. Wenzel)  ECCO.
COMPUTER MODELS IN BIOLOGY Bernie Roitberg and Greg Baker.
CF-3 Bank Hapoalim Jun-2001 Zvi Wiener Computational Finance.
Nonlinear and Non-Gaussian Estimation with A Focus on Particle Filters Prasanth Jeevan Mary Knox May 12, 2006.
Optimal Bandwidth Selection for MLS Surfaces
The Calibration Process
458 Fitting models to data – III (More on Maximum Likelihood Estimation) Fish 458, Lecture 10.
Bayesian Analysis of Dose-Response Calibration Curves Bahman Shafii William J. Price Statistical Programs College of Agricultural and Life Sciences University.
Model Checking in the Proportional Hazard model
Optimization of thermal processes2007/2008 Optimization of thermal processes Maciej Marek Czestochowa University of Technology Institute of Thermal Machinery.
© 2008 McGraw-Hill Higher Education The Statistical Imagination Chapter 9. Hypothesis Testing I: The Six Steps of Statistical Inference.
Copyright © Cengage Learning. All rights reserved. 8 Tests of Hypotheses Based on a Single Sample.
Introduction to Adaptive Digital Filters Algorithms
Combinatorial Reconstruction of Sibling Relationships in Absence of Parental Data Tanya Y Berger-Wolf (DIMACS and UIC CS) Bhaskar DasGupta (UIC CS) Wanpracha.
Particle Filtering in Network Tomography
ChE 452 Lecture 24 Reactions As Collisions 1. According To Collision Theory 2 (Equation 7.10)
SIS Sequential Importance Sampling Advanced Methods In Simulation Winter 2009 Presented by: Chen Bukay, Ella Pemov, Amit Dvash.
1 HMM - Part 2 Review of the last lecture The EM algorithm Continuous density HMM.
A Modified Meta-controlled Boltzmann Machine Tran Duc Minh, Le Hai Khoi (*), Junzo Watada (**), Teruyuki Watanabe (***) (*) Institute Of Information Technology-Viet.
Probabilistic Robotics: Monte Carlo Localization
Stochastic Linear Programming by Series of Monte-Carlo Estimators Leonidas SAKALAUSKAS Institute of Mathematics&Informatics Vilnius, Lithuania
Statistics for Managers Using Microsoft Excel, 5e © 2008 Pearson Prentice-Hall, Inc.Chap 8-1 Statistics for Managers Using Microsoft® Excel 5th Edition.
Particle Filters for Shape Correspondence Presenter: Jingting Zeng.
Heuristic Optimization Methods
VI. Evaluate Model Fit Basic questions that modelers must address are: How well does the model fit the data? Do changes to a model, such as reparameterization,
1 IE 607 Heuristic Optimization Particle Swarm Optimization.
LOGISTIC REGRESSION David Kauchak CS451 – Fall 2013.
Evolving Virtual Creatures & Evolving 3D Morphology and Behavior by Competition Papers by Karl Sims Presented by Sarah Waziruddin.
Sophie RICCI CALTECH/JPL Post-doc Advisor : Ichiro Fukumori The diabatic errors in the formulation of the data assimilation Kalman Filter/Smoother system.
Integrating archival tag data into stock assessment models.
Thursday, May 9 Heuristic Search: methods for solving difficult optimization problems Handouts: Lecture Notes See the introduction to the paper.
Lecture 4: Statistics Review II Date: 9/5/02  Hypothesis tests: power  Estimation: likelihood, moment estimation, least square  Statistical properties.
Mobile Robot Localization (ch. 7)
How will storage change and discharge be estimated from SWOT? Michael Durand, Doug Alsdorf, Ernesto Rodríguez, Kostas Andreadis, Elizabeth Clark, Dennis.
Copyright © Cengage Learning. All rights reserved. 13 Linear Correlation and Regression Analysis.
1 Introduction to Statistics − Day 4 Glen Cowan Lecture 1 Probability Random variables, probability densities, etc. Lecture 2 Brief catalogue of probability.
Chapter 7 Estimates and Sample Sizes 7-1 Overview 7-2 Estimating a Population Proportion 7-3 Estimating a Population Mean: σ Known 7-4 Estimating a Population.
Evolving RBF Networks via GP for Estimating Fitness Values using Surrogate Models Ahmed Kattan Edgar Galvan.
A comparative approach for gene network inference using time-series gene expression data Guillaume Bourque* and David Sankoff *Centre de Recherches Mathématiques,
Paging Area Optimization Based on Interval Estimation in Wireless Personal Communication Networks By Z. Lei, C. U. Saraydar and N. B. Mandayam.
2005 Unbinned Point Source Analysis Update Jim Braun IceCube Fall 2006 Collaboration Meeting.
ESTIMATION METHODS We know how to calculate confidence intervals for estimates of  and  2 Now, we need procedures to calculate  and  2, themselves.
Active Walker Model for Bacterial Colonies: Pattern Formation and Growth Competition Shane Stafford Yan Li.
Meta-controlled Boltzmann Machine toward Accelerating the Computation Tran Duc Minh (*), Junzo Watada (**) (*) Institute Of Information Technology-Viet.
Swarm Intelligence. Content Overview Swarm Particle Optimization (PSO) – Example Ant Colony Optimization (ACO)
Heuristic Optimization Methods
12. Principles of Parameter Estimation
Particle Swarm Optimization
Boosting and Additive Trees
Some Design Recommendations For ASAP Studies
More on HW 2 (due Jan 26) Again, it must be in Python 2.7.
12. Principles of Parameter Estimation
Population Methods.
Presentation transcript:

Geolocation of Icelandic Cod using a modified Particle Filter Method David Brickman Vilhjamur Thorsteinsson

What does one do when …

Note that the T simulation is good, but the recapture estimate is way off Note that track goes into deep water – not considered likely for Icelandic Cod Varying parameters improves results but not by that much. The best that a “standard” particle filter can do is DST recap position model recap position DST tag position model simulation 600m 200m

Why does this occur?? T field around Iceland is ~ parabolic so that particles drifting from tag location, and trying to follow T data, have 2 possible directions to choose. Temperature field ~ parabolic Climatological September T at 100m Aside: T field for this study comes from a state-of-the-art circulation model for the Iceland region developed by Kai Logemann (Logemann and Harms Ocean Sci., 2, 291–304, 2006)

OPTIONS 1. Accept that this is the best that the PF method can do and Do nothing Hide these results (~5 out of 27) 2. See whether modifications to the PF method can produce better simulations

The Data: 27 useable DSTs Example of tag being inserted into cod fish (from Star-Oddi website)

Example of DST data

Movement model: Where x n = (lon,lat) position at time n = the “state” V = (max) swim velocity = model parameter U = random # from uniform distribution dx = change in (lon,lat) position dt = timestep “Standard” Particle Filter (PF-1) (Andersen et al. 2007, CJFAS 64: ) dxdx Vdt xnxn x n+1 particle z-level = DST z-level Particles start at the initial tagging position, and evolve according to a

Observation model: Where y n = observation at time n (i.e. temperature) from the DST NB: last time includes the “recapture” observation  = error g(x) is the model temperature field at x derived from a numerical circulation model

Error Model -- Particle Filter Weight: Standard assumptions for a SIR filter yield: The probability of the observation given the state is and following Andersen et al. (and others):

Particle Filter: “PF-1” At t=0 NP particles are seeded at the (known) DST tagging position Each particle evolves according to the movement model (A) At each timestep evaluate P  particle filter weights w (B) Sample with replacement NP particles from w, preferentially choosing those with higher probability (i.e lower error). Use the standard SIR cumulative distribution method. (C) Propagate these particles to the next step Repeat A-C Continue to end of series, at which time the recapture position is an (important) observation to be incorporated into P. NB: no backward smoothing procedure coded.

Example of a Good Result Standard PF PF-1 However, note that offshelf drift is not considered biologically realistic

Modifications to standard PF Two modifications added: “Attractor” function: To increase the influence of the final (recapture -- R) position, a time-dependent term was added to the error model: distance from recapture position factor that increases as final time is approached Allows a future observation to influence present state Adds 2 parameters: time0 and  a

Interpretation of Attractor term Consider 2 particles returning the same T (i.e. T-error) – late in the simulation The estimate reported by particle 2 is considered more likely because it is closer to the recapture position. 1 2 recap position

Demersal error term: Intended to correct the tendancy for particles to follow increasing temperatures by drifting southward Observed in many simulations but considered biophysically unlikely. where z n i is the depth of the i-th particle at time n (= DST depth) and z btm (x n i ) is the model bottom depth at location x n i  d is a vertical scale parameter

Interpretation of Demersal term Assumes that the school of fish are clustered within  d of the bottom and penalizes those fish that do not fit into this “demersal” vertical distribution. Consider 2 particles at the same depth, reporting the same T the estimate reported by particle 2 considered more likely as that fish is exhibiting a more demersal behavior. Action ~ negative diffusion 1 2

New terms incorporated in an error distribution (at every timestep, for each particle): E =  {T-error + attractor-term + demersal-term} i.e. additive error distribution of un-normalized error terms, sampled using a SIR-type procedure; New Error Model Preferentially choose particles with lowest error

How to think about this -- Heuristically For this type of problem (i.e. DST) the backward smoothing procedure is essential as it is the way that the recapture observation influences the solution: Up to recap obs, PF yields optimal “local” solution Use of backward smoothing produces optimal “global” solution. E-distribution “attempts” to solve the global problem in one pass through the data.

Regarding E -- Consider minimizing a likelihood function over all observations: BTW: solved L(y|  ) for optimal parameters using a Direct Search algorithm. DS algorithm: see Kolda et al. 2003, SIAM V.46, no.3, pp

Note that the Demersal term could be incorporated into the Movement Model by including bathymetry: shallow deep shallow deep dxdx Vdt xnxn x n+1 Present model Model using bathymetry dxdx Vdt xnxn x n+1

Results

Addition of Attractor function only (PF-2) Cf: no attractor

Addition of Attractor function plus Demersal term (PF-4) Cf: attractor only

Comparison of PF-1 versus PF-4 ( NB: Different DST)

Summary / Conclusion Standard PF seen to perform poorly on a number of DSTs PF method modified by adding: Attractor term that “sucked” particles toward the recapture position allows future data to influence present result Demersal term that favoured particles that adhered to a “gadoid-type” behavior keeps particles onshelf Attractor + demersal terms can be considered to be rules or behaviors imposed on the particles. Result likely depends on Temperature field: demersal term may not be necessary

Forcing better biological behavior (addition of demersal term) resulted in poorer simulation of temperature timeseries. i.e. quantitatively WORSE results “Best” result is subjective OR Modified PFs performed better than standard PF, especially on difficult DSTs. However, When signal processing theory meets fisheries biology adjustments may have to be made