Developers Meeting Notes 3/2/12

No replies
rgore's picture
Offline
Joined: 01/27/2011

At developers meeting on 3/2 we discussed the following:

  • The group discussed several different strategies for Mike Hunter to employ when debugging his LISREL-related code additions to OpenMx. Specifically, it is very difficult to debug R code on Windows because of a lack of gdb support. Mike is going to install a virtual Linux distribution via Virtual Box and attempt to debug through Linux.
  • The group discussed the performance of the implemented analytic/smart gradients. The gradients save estimations but not iterations in the optimization. Michael Spiegel found that depite offering theoretical performance improvements the current implementation results in 5x slowdown on the traditional OpenMx test suite programs. Tim has made some changes to the code to reduce the number of matrix multiplications in the computation of the analytic gradient but there still appears to be a 4x slowdown. Tim is going to continue to look into this.
  • After a productive meeting with Tim, Ross has been able to start to make progress in implementing the solnp and subnp algorithms in the R package Rsolnp into R. For now he is avoiding dealing with function pointers and just hardcoding in several of the benchmark examples. Once these examples are working, function pointers will be employed so that the test problems can become parameterized.
  • The group discussed the differences that can exist in estimates resulting from the choice to use raw data (FIML) vs. covariance data (ML) when the sample size is small. This issue was raised and discussed in detail in this thread. The group was opposed to changing the current behavior of FIML or ML but discussed the possibility of adding new functions that would adjust for the differences in the scaling factors between the two approaches so that they would produce the same results for a given model.
  • The bulk of the meeting was spent discussing Least Squares user interface proposals. The goal to implement Weighted Least Squares and Diagonal Least Squares will cause additions and modifications to the Least Squares user interface, and possibly several other user interfaces. The group will continue discussing the ramifications of several different proposals put forth by Ryne and Tim in future meetings and via email.