• | Posterior maximization algorithm: You can choose between Christopher Sims' csminwel, Marco Ratto's newrat, Dynare's gmhmaxlik, and matlab's fminunc optimization routines and to target either the transformed parameters or the original parameters for the optimization. In the case of the transformed parameters, YADA first transforms the original parameters, whose supported is bounded, onto the real line and then take the Jacobian of the transformation for the prior into account in the computation of the log posterior. In the case of the original parameters, YADA does not transform the parameters, but instead sets a very high penalty to parameter values that are outside the bounds of the prior. The fminunc alternative is only available if the user has the YADA specific version of fminunc. To get access to YADA specific version of fminunc, the user needs to edit their own fminunc files according to the details laid out in a set of diff-files. These diff-files are available for download from the YADA website. Instructions on required name-changes and the specific location(s) that the new files should have within the YADA directory structure can be obtained by downloading the diff-files package, while pointers on how to make use of the diff-files are provided in the FAQ. |
• | Tolerance level: Choices of tolerance level for the maximization routine range from 0.01 to 0.00000000001. |
• | Maximum number of iterations: Lets you select the the upper limit for the number of iterations. Integer values between 1,000 and 1,000,000,000 are supported! |
• | Method for initializing inverse Hessian: The initial value of the inverse Hessian when running the csminwel and newrat optimization routine is times 0.001 times the identity matrix. This option allows the user to select different initial values for this matrix: they range from 10 times the identity to 0.00001 times the identity. |
• | Maximum number of extra csminwel runs: When csminwel is used for optimization the convergence criterior is given by the change in the value of the log posterior. If the gradient vector is larger than the selected tolerance level, the user may wish to rerun csminwel with the latest value for the estimated parameters as the initial value but with the inverse Hessian being reinitialized. This control allows the user to select the maximum number of such extra runs of csminwel. The default is 0 while the maximum is 20. |
• | Grid width when checking the optimum [width / number of points]: YADA sets up a grid around the posterior mode when checking the properties of the log posterior around that point. The possible width of the grid is 1 to 5 times the standard deviation of the parameter. The standard deviation is here given by the square root of the corresponding diagonal element of the inverse Hessian that is supplied by the optimization routine. This grid is always defined from the transformed parameters. While the control to the left lets you influence the upper and lower limits of the grid, the number of points in the grid is determined by your choice in the control to the right. Integer values between 10 and 100 are supported. |

Additional Information
• | A more detailed description about parameter transformation is given in Section 6 of the YADA Manual. |
• | A more detailed description about posterior mode estimation is found in Section 7 of the YADA Manual. |
Page url:
http://www.texlips.net/yada/help/index.html?optimization.htm