and Simulation, Bauhaus Summer School 2012 Model Validation , -, 3-5 2011
Weimar, Germany August 17, 2012
Project 5:
Calibration of Numerical Models in Civil Engineering Applications
Adrienn Tth, Hungary Dafni Skiada, Greece Jana Kadrov, Czech Republic Luciane Marcela Filizola de Oliveira, Brazil Savvas Papadopoulos, Greece ahin Bankir, Turkey Vesile Hatun Akansel, Turkey Coordinator: Dr. Tom Lahmer
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Introduction
Models are only valuable if one knows the model parameters exactly We obtained parameters values based on measured data model calibration (inverse modeling)
General questions to answer through optimization:
Which parameters provide the best fit? Can these parameters be identified uniquely? What is the confidence in these parameters? Are there enough data?
Does my model fit to the data? Validation and selection of models!
How do error components in the data influence my optimization results? How to judge about the quality of my model and the data after the calibration?
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Task: Harmonic oscillator
The motion of an unforced harmonic oscillator satisfies the initial value problem
where is the mass, c the damping and k the stiffness of the system
Task:
Intro
Identification of , c and k from transient deflection data
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Gradient Method
Features of the Gradient Method Minimizing differentiable function f(x) Guessing starting point and choose the direction in which decreases most quickly (in the negative gradient direction)
Advantages of the Gradient Method
Every iteration is inexpensive Does not require second derivatives
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
BFGS (Broyden, Fletcher, Goldfarb, Shanno) Method
Features of the BFGS Method It approximates Newtons Method (Quasi-Newton Method) The Hessian matrix is approximated
Advantages of the BFGS Method
BFGS has proven good performance even for non-smooth optimizations It is more efficient than Gradient Method
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods Bauhaus Summer School 2012 Model Validation and Simulation, in Structures Dynamics and Earthquake Engineering
Comparison Gradient Method x BFGS Method
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods Bauhaus Summer School 2012 Model Validation and Simulation, in Structures Dynamics and Earthquake Engineering
Comparison Gradient Method x BFGS Method
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods Bauhaus Summer School 2012 Model Validation and Simulation, in Structures Dynamics and Earthquake Engineering
Comparison Gradient Method x BFGS Method
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods Bauhaus Summer School 2012 Model Validation and Simulation, in Structures Dynamics and Earthquake Engineering
The Nelder-Mead Method
x1, x2, xn f(x1) f(x2) f(x3) Reflect: xr
f(x1) f(xr) f(x2) x1,f(x1) xcc,f(xcc) x2,f(x2) x3,f(x3)xn+1,f(xn+1)
f(xr)< f(x1) Expand: xe f(xe)< f(xr) xr
xe
f(x2) f(xr) f(x3) Contract outside: xc f(xc) f(xr) xr xc Shrink Contract inside: xcc f(xcc)< f(x3) xcc
xc,f(xc)
xr,f(xr)
xe,f(xe) Terminate the iteration
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
The searching process
Initial guesses a. K=2.6 & c=0.5 b. K=3.0 & c=0.7 c. K=3.1 & c=1.2
These are the 3 first vertices of the triangle
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
The Nelder-Mead Method
Fitted data match the measured data There is no noise in the signal
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Particles Swarm Optimiser The PSO is a stochastic, population-based computer algorithm based on swarm intelligence. It is a direct, derivative-free global search algorithm.
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Particles Swarm Optimiser
Flow chart depicting the General PSO Algorithm:
Start
Initialize particles with random position and velocity vectors.
Loop until all particles exhaust For each particles position (p) evaluate fitness If fitness(p) better than fitness(pbest) then pbest= p Set best of pBests as gBest Update particles velocity and position Stop: giving gBest, optimal solution.
Intro Gradient methods Nelder-Mead Particle Swarm Optimiser Bayesian updating Conclusions
Loop until max iter
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Particles Swarm Optimiser
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Particles Swarm Optimiser
Results
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Particles Swarm Optimiser
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Bayesian Updating Updating of parameters/model based on measurements
Stochastic method
parameters are identified with their probability distributions Likelihood function gives the probability of observing measured data with
chosen parameters values
Different methods how to calculate the likelihood and probabilities: Kalman filter
Method acc. to Bazant
Gibbs sampling Metropolis-Hastings algorithm
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Bayesian Updating
Metropolis-Hastings algorithm
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Conclusion
Can these parameters be identified uniquely?
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Conclusion
Comparison of different algorithms
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
3rd International Conference on Computational Methods in Structures Dynamics and Earthquake Engineering Model Validation and Simulation, Bauhaus Summer School 2012
Thank you for your attention.
Intro
Gradient methods
Nelder-Mead
Particle Swarm Optimiser
Bayesian updating
Conclusions
