Approach to model calibration in TreeAge Pro
I am interested in the community’s approach to model calibration in the TreeAge environment?
If modelers are attempting to estimate missing model parameters or else to adjust model parameters through calibration to observed data, and the number of parameters to be estimated or adjusted is large, how have they searched the parameter space?
I have developed techniques using Latin hypercube sampling within TreeAge as well as simulated annealing by linking TA with Excel. I’ve also developed Python user functions to analyze Kaplan-Meier survival curves internally within TA and to use these and the calibration targets to calculate Euclidian distance GOF metrics.
How have others done this?
Has anyone implemented a Nelder-Mead algorithm or Pareto optimization in TA? Which GOF metrics have they used?
Comments
I have implemented model calibration framework using Scientific Python (SciPy and NumPy) and TreeAge Pro. The SciPy optimization library includes about 10 different algorithm and Nelder-Mead is one of the available algorithms. Simulated annealing has been replaced by a basin hopping algorithm that has a similar "temperature" parameter. With some experiments of calibrating a simple model with 6 degrees of freedom (finding rates and shapes of 3 different Weibull distributed transitions), Nelder-Mead algorithm seems to be more efficient than basinhopping.
For more details feel free to reach out to me at support@treeage.com
Regards,
Al
Hello, David. I was looking for information about using latin hypercube sampling and other methods to reduce computational burden with TreeAge and came across your post.
May I reach out to you with regards to how you used this sampling method within TreeAge?
Regards,
Madhu.
For sure Madhu
I'm at david.naimark@sunnybrook.ca
Model calibration example using Python and detail information about installation and running the Python calibration session using Nelder-Mead algorithm can be found at our website by following this link.
Please sign in to leave a comment.