At the Developers' Meeting on 4/13/12 we discussed the following:
- Came to consensus on implementing the implementation of changes to the interface to accomodate numerous different types of fit functions. A writeup of the proposed changes will be made available to OpenMx users. These changes have been devised to maximize future fit functionality and minimize changes to the existing interface.
- Tim Brick has identified several modifications that need to be made in the implementation of analytical gradients to achieve speedup.
- The current implementation of Csolnp (The C translation of the solnp optimizer) can solve the Powell Problem benchmark. Current testing for other benchmarks is still needed as well as the error checking for the inputs.
- Dan Hackett has implemented a latent model with many variables and many manifests. PPML, which is currently implementing in the front end, is not achieving the expected Speedup. The group believes the bottleneck is in R and PPML needs to be implemented in the back end for additional speed up to be achieved.
- The group discussed several algorithmic techniques to speedup fitting models.
- Mike Hunter has continued to make progress on implementing LISREL models. He has implemented error checking for the front end and created a test case that exercises each error. Compared to RAM models LISREL is only 2-10% slower. A user guide is upcoming but interested users should check out the code from the trunk and begin to experiment with it.