Dear all. I am trying to maximize a complex log likelihood function with respect to 10 parameters using the method "L-BFGS-B" implemented in the optim procedure . The algorithm that I have written always converges, but I have got different solutions running the algorithm many times on the same dataset. I suspect that the log likelihood is flat.... The results usually differ of a quantity that is small (+- 0.04). But, given that some of the parameters can vary from -1 to 1, I am not satisfied with the variability in the solutions and I do not know how to choose a solution among the results that I get. I have tried to iterate the optim procedure, using as initial values the results of the previous step to see if the algorithm does rich the convergence.But it does not, after 1000 iterations. I would like to summarise the results of the maximization procedure at the different iterations as an estimator of the unknown parameters, for instance, as a kind of MC average. Does anybody have any expercience in such a theme? I would really appreciate comments or ideas! It is very important! Many thanks, Annarita Annarita Roscino Department of Statistical Sciences University of Bari tel. 00390805049353 email: aroscino at dss.uniba.it
Dear all. I am trying to maximize a complex log likelihood function with respect to 10 parameters using the method "L-BFGS-B" implemented in the optim procedure . The algorithm that I have written always converges, but I have got different solutions running the algorithm many times on the same dataset. I suspect that the log likelihood is flat.... The results usually differ of a quantity that is small (+- 0.04). But, given that some of the parameters can vary from -1 to 1, I am not satisfied with the variability in the solutions and I do not know how to choose a solution among the results that I get. I have tried to iterate the optim procedure, using as initial values the results of the previous step to see if the algorithm does rich the convergence.But it does not, after 1000 iterations. I would like to summarise the results of the maximization procedure at the different iterations as an estimator of the unknown parameters, for instance, as a kind of MC average. Does anybody have any expercience in such a theme? I would really appreciate comments or ideas! It is very important! Many thanks, Annarita Annarita Roscino Department of Statistical Sciences University of Bari tel. 00390805049353 email: aroscino at dss.uniba.it
Apologies for the multiple postings. Dear all. I am trying to maximize a complex log likelihood function with respect to 10 parameters using the method "L-BFGS-B" implemented in the optim procedure . The algorithm that I have written always converges, but I have got different solutions running the algorithm many times on the same dataset. I suspect that the log likelihood is flat.... The results usually differ of a quantity that is small (+- 0.04). But, given that some of the parameters can vary from -1 to 1, I am not satisfied with the variability in the solutions and I do not know how to choose a solution among the results that I get. I have tried to iterate the optim procedure, using as initial values the results of the previous step to see if the algorithm does rich the convergence.But it does not, after 1000 iterations. I would like to summarise the results of the maximization procedure at the different iterations as an estimator of the unknown parameters, for instance, as a kind of MC average. Does anybody have any expercience in such a theme? I would really appreciate comments or ideas! It is very important! Many thanks, Annarita Annarita Roscino Department of Statistical Sciences University of Bari tel. 00390805049353 email: aroscino at dss.uniba.it