At developers meeting on 10/14 we discussed the following:
- Dan Hackett is investigating some issues with the standard errors in PPML. The standard errors do not match OpenMx for some larger container models. Dan is investigating this further. We believe this is because the standard errors are unstable. Thus even the same algorithms are capable of producing different standard errors because the estimates are moving around in space so much.
- The group discussed the efficiency improvements offered by the thread level parallelism. Specifically the almost 4x improvement for startsTestMissing.R contrasting to no improvement for startsTestNoMissing was discussed. There is no speedup for startsNoMissing because there were already a number of optimizations in OpenMx to improve efficiency in this case. However, in the case of startsTestMissing.R there not any previous optimizations. startsTestMissing.R makes numerous calls to the objective function where the hessian parallelization offers speedup.
- The group discussed an identified bug in the implementation of thread level parallelism. The bug is related to the state duplication when summing mxAlgebras. The group is looking into this.
- The group discussed how FIML parallelization would work in concert with the hessian thread-level parallelization.
- The group discussed two bugs in OpenMx that had been identified via the forums.