Developer's Meeting Notes 7/27/12

No replies
AdminRoss's picture
Offline
Joined: 03/07/2011

At the Developers Meeting on 7/27/12 we discussed the following:

  • The issues with 'make test' discovered by Mike Hunter in last week's meeting have been resolved. The signatures for the MxMLObjective and MxAlgebraObjective did not contain a 'labelsData' argument in their genericObjFunConvert methods.
  • Michael Spiegel has been implementing dependency tracking of free parameters. The first goal is to get the dependency tracking working for single-threaded models first and then for multi-threaded cases.
  • The group discussed the progress of csolnp and where and how it fits in to OpenMx and R. The current vision is that csolnp will be able to be called directly from R using the 'inline' package. The group also discussed the possibility of using csolnp in a multi-threaded fashion by eliminating the use of global variables in the current implementation. The source code is currently being debugged for memory leaks and refactored to conform with standards in the BLAS library (and to be able to return RObjects in the 'inline' package).
  • The group continued to discuss the implementation and interpretation of confidence intervals when a boundary condition is reached during optimization and the solution lies outside the boundary. The group has decided to report the confidence intervals and tell the user in LARGE LETTERS that multiple interpretations/answers of the confidence intervals are possible and they get to pick which interpretation they want.
  • Joshua Pritkin progress especially with employing EM algorithms and how their optimization fits into OpenMx. This led to a discussion about several customization features that would be nice to have for the OpenMx optimzer (npsol, csolnp). In particular two features were identified:
    • The ability to divide a model into parents and children. Where the parent optimization would stop and a new child optimization would start that the parent would wait for and the once complete would employ once the parent optimization was restarted. The group hypothesized this would a general way to consider swarm optimization.
    • The ability to have a moving target within the optimization process. The group speculated that there was a number of novel contributions that could be made to enable this feature, however, a literature survey should be done.
  • The group discussed the state of funding opportunities for OpenMx and other projects.
  • The group discussed the state of 'omxBrownie' which is currently checked into the source code trunk.
  • Mike Hunter reaffirmed his goal to implement state space models for OpenMx. Currently there is lack of standardization for the specification of these models. As a result, Mike will be surveying the literature to identify which specification OpenMx will employ and popularize.