FAQ
-
Shall I write the equations of the model in a given order?
No, you can write the equations in the order you prefer.
Remember, in Eas_stimate you can enter your model the way you would write it on a paper pad. Back to the list
-
Can I change the symbols for dependent and independent variables?
No, you can't. The symbols x, y, w, s, m, z and q are reserved. An error message is visualised if you do not comply with this notation.
Back to the list
-
Can I use stars, primes, carets, etc. to identify variables?
No, you can only use subscripts to identify variables.
Back to the list
-
Is the definition of variables case-sensitive?
Yes, small and capital letters are different variables.
Back to the list
-
Why can't I decide on the type of objective function to use?
The most suitable likelihood function to be maximised is selected automatically by Eas_stimate, according to the information contained in the model and in the variance.
Back to the list
-
I don't have reliable values for the variance matrix. Shall I use the option that makes it possible to estimate it?
Yes, you can. However, if your starting parameters values are not close enough to the true values, the convergence can be slow or even unattainable.
You'd better make an educated guess on the variance matrix and improve your parameters first.
Back to the list
-
What is a-priori information?
You can have some information on the values of the parameters either from previous experimental campaigns or from the literature.
Back to the list
-
What is a run?
In differential models you frequently have experiments with different initial conditions, time sampling and independent variables. Each of them is a run.
Back to the list
-
What if I have more than 20 runs?
Use the first 20 runs to estimate the parameters. Improve the estimates by running Eas_stimate with the remaining runs after the previously estimated parameters have been entered as a-priori information.
Back to the list
-
Does Eas_stimate identify outliers?
No, it does not. The user can eliminate the data that he suspetcs to be outliers after visual inspection of the results.
Back to the list
-
Can I use non linear inequality constraints on either parameters or algebraic variables m?
No, you can't. Non linear inequality constraints are indeed very unusual in physical models. Additionally, they would make convergence difficult.
Furthermore, remember: algebraic variables and parameters of error-in-variables models can be constrained only using simple bounds.
Back to the list
-
How many missing data points can there be?
In principle there is no limitation. However, remember that any missing data point requires additional computations. This may jeopardise and/or slow down convergence.
Back to the list
-
How many iterations should be allowed?
It depends on the problem. Start with the default value and increase the number of iterations according to he progress in the reduction of the objective function.
Back to the list
-
Why should I treat linear parameters differently from the non linear ones in algebraic models?
In some cases separating linear from non linear parameters can improve the convergence. In others it does not. It's better you experiment with both methods.
Back to the list
-
Why should I change the Fortran modules generated by Eas_stimate?
Don't change the Fortran code if you were able to enter your mathematical model in the main window. However, some models may require manual adjustments: for instance discontinuities (depending on the values of a variable). Follow the instructions contained in the comments of each Fortran module provided by the Fortran editing windows.
Back to the list
-
Which optimisation method should I use?
Try the Gauss-Newton method first. If the residuals (difference of computed to experimental values) are large, try the Quasi-Newton method.
Back to the list
-
Why should I worry about the inversion of the Hessian matrix?
In highly sensitive models (or if your tentative parameters are very far from the correct ones) the Hessian matrix of the likelihood function can be nearly singular. While the Marquardt method (which is considerably faster)
is adequate in most cases, you may need to use the singular value decomposition method in very sensitive problems.
Back to the list
-
The procedure for estimating parameters in a parabolic partial differential equation is very inefficient. How can I speed it up?
PDE models should be used only after more approximate (ODE) models have provided good estimates for the parameters. A limited number of parameters (two or three) can be fine-tuned using the more complete PDE model.
Back to the list
-
When should I make use of the "preliminary steps"option?
An efficient random search is carried out prior to the full gradient-based maximisation of the likelihood function, if you choose this option. Global (as opposite to the local optimum located by gradient-based algorithms) may be detected.
However, there is no guarantee that this actually happens.
Back to the list
-
Why should I want to introduce addtional grid points in the plot of the differential variables vs. time?
The graphical plot is obtained by interpolating numerical values. If a discontinuity is included between two consecutive grid points, it is likely to be smoothed out. By choosing to have this point in the graph, the real numerical value is plotted.
Back to the list
-
Do I interfere with the optimisation procedure, if I interrupt it to examine numerical values of the graphs being plotted?
No, you don't. Simply press on the "Resume"command to get the procedure to continue its optimisation task.
Back to the list
-
How can I select one of the plots during a pause (or at the end of the optimisation)?
Simply click on it. Its border will turn red.
Back to the list
-
Can I insert the text of the model into Office applications?
You can import the "sym"file as an "rtf"application. However, a few coded lines related to the additional information on the model (variance, starting values, etc.) will appear at the end of it.
Back to the list
-
How can I get the solution of an estimation problem, if I don't want to buy the package?
You can download the demo programme. It is a fully functional form, but for the fact that it does not show the values of the parameters, neither does it write the results to the output file.
However, looking at the fit, you can verify if convergence has been attained. If it has, send us model and data and upon payment of a lump sum of 50 US$, you will receive the full output file.
If convergence has not been attained and you want us to obtain it for you, contact us and we shall let you know if and at what price we can work it out for you.
Back to the list
-
Who is presently using Eas_stimate?
In addition to several academic establishments, Eas_stimate is available for use to all companies of the ENI Group (17th largest group worldwide in 2009).
Back to the list
|