Convergence of the full_analysis.py script.Convergence of the full_analysis.py script.Hi, As optimisation should converge very quickly when the differences between iterations is at such low significant figures, I'm unsure as to what is causing the problem. For the spheroidal and ellipsoidal diffusion tensors convergence should occur between 10-15 iterations, and maximally after 20. The reason could be because the script is trapped jumping between the same two models for each alternate iteration. I have a feeling though that this is an issue caused by over-fitting and the parameter values are just drifting aimlessly through the parameter space. Actually, thinking about this, I can almost guarantee that that is what is happening. The data at only a single field strength (https://mail.gna.org/public/relax-users/2007-04/msg00006.html) is insufficient for the full_analysis.py script. Sorry for not responding earlier. Cheers, Edward On Mon, 2007-04-30 at 15:11 +0100, Clare-Louise Evans wrote:
Dear Edward, Sorry for the thread hijack, but I noticed in the PS section of the quote Hongyan has included below your comment regarding convergence. I am having problems with convergence at the moment when running the full_analysis.py script. I have run the optimisation of the diffusion models on the same PC. For the oblate, prolate and ellipsoid models I have failed to reach convergence after 30+ rounds. The sphere model converged within 5 rounds. However, when I look at the output the only difference between the chi-squared and other parameter values is in the 10th and later decimal places. Surely, with having such small differences between the values they can be considered to have converged. However, in your comment you state they have to be identical. If this really is the case then I'm not sure how to proceed given that I'm failing to reach convergence on these 3 models? Apologies if this ties in to the question I asked yesterday regarding full_analysis.py. Kind regards Clare Hongyan Li wrote:Dear Edward, Thanks for your early suggestion regarding different model selections. I have been busy on other stuff and only recently got time to focus on this subject again. I have run all the diffusion models e.g. isotropic, prolate, oblate and ellipsoid and model selections were made within each model (in aic directory). I am not sure how to write a script to select different models using 'model_selection()'. Could you please do me a favor? Your help is highly appreciated! Best wishes, Hongyan Quoting Edward d'Auvergne <edward.dauvergne@xxxxxxxxx>:Hi, To compare the results what you need to employ is a technique from the statistical field of model selection. The spherical diffusion (isotropic) + all model-free models of all selected residues is one single mathematical model. The prolate and oblate spheroids (prolate and oblate axially symmetric anisotropic diffusion tensors) + all model-free models, and the ellipsoid (fully anisotropic or three different eigenvalues) + all model-free models, are three additional mathematical models. Therefore to compare these four different models you need to select the model which best represents your relaxation data. These models are, however, not nested and therefore cannot be compared using ANOVA F-tests! Firstly the three types of diffusion tensor are not nested (there is a reference from Dominique Marion's group in which they say ANOVA statistics cannot be used but I can't find it at the moment (although it shouldn't be too hard to track down, it's related to Tensor)). Secondly the model-free models selected will be different between the four models. Hence chi-squared and F-tests cannot be used. A useful reference (I'm not at all biased ;) for this problem is my paper d'Auvergne, E. J. and Gooley, P. R. (2003), (see http://www.nmr-relax.com/refs.html for the full reference). On page 37 at the end of that paper I discuss how AIC model selection is perfect for selecting between these non-nested models. The AIC criterion is still AIC = chi2 + 2k, however chi2 is the minimised chi-squared value for the complete model and k is the sum of the number of diffusion parameters and number of model-free parameters for all spin systems. BIC model selection is likely to work quite well as well. If you have four runs, one for each of the diffusion models, then the relax user function 'model_selection()' is designed to select between these models. I hope this helps and hasn't been too biased. Cheers, Edward P.S. Prior to model selection between the diffusion models, the diffusion models must have fully converged. Multiple iterations of optimisation of the model-free models, AIC model selection, and optimisation of all parameters together (diffusion tensor + model-free parameter of all residues) must be executed. Convergence is when two iterations possess identical chi-squared values, identical model-free models, and identical parameter values. On 3/2/07, Hongyan Li <hylichem@xxxxxxxxxxxx> wrote:Dear relax users, I have managed to use Relax to run my dynamics data by both isotropic and axial-oblate models. My qestion is how to compare the results, bychi-square??what is the criteria to make judgment that which residues is betersimulatedwith which model? Thanks for your kind help! Best wishes, Hongyan Dr. Hongyan Li Department of Chemistry The University of Hong Kong Pokfulam Road Hong Kong _______________________________________________ relax (http://nmr-relax.com) This is the relax-users mailing list relax-users@xxxxxxx To unsubscribe from this list, get a password reminder, or change your subscription options, visit the list information page at https://mail.gna.org/listinfo/relax-usersDr. Hongyan Li Department of Chemistry The University of Hong Kong Pokfulam Road Hong Kong _______________________________________________ relax (http://nmr-relax.com) This is the relax-users mailing list relax-users@xxxxxxx To unsubscribe from this list, get a password reminder, or change your subscription options, visit the list information page at https://mail.gna.org/listinfo/relax-usersThis message has been checked for viruses but the contents of an attachment may still contain software viruses, which could damage your computer system: you are advised to perform your own checks. Email communications with the University of Nottingham may be monitored as permitted by UK legislation. _______________________________________________ relax (http://nmr-relax.com) This is the relax-users mailing list relax-users@xxxxxxx To unsubscribe from this list, get a password reminder, or change your subscription options, visit the list information page at https://mail.gna.org/listinfo/relax-users