At the Developers Meeting on 8/24/12 we discussed the following:
- Steve discussed the group's original approach to multi-level modeling. This approach generated some inefficiency when scaling. In particular, larger numbers of observations caused a quadratic time increase. However, within the last week Steve and Tim have discovered a strategy that groups observations and covariances by pairs which results in only a linear time increase. In short, the current approach to multi-level modeling takes tall matrices and turns them into wide and then works on the matrices in the wide matrices.
- The group also discussed the topic for an upcoming paper that discusses an information theoretic approach to determining the amount of information gained from a new measurement/prediction by the optimizer. In particular, what does this measurement suggest about a new dynamic approach to optimization that is distributed in a 'Middle' type of way. Mike Hunter suggested looking into the Kullback-Leibler Divergence to measure this.
- The group discussed how this paper would help potential funding opportunities. The group also discussed potential funding opportunities for this work and OpenMX
- Tim Brick discussed and distributed the new proposal for mxFit and mxExpectation.
- Rhyne discussed his progress with WLS and a current test case that is reporting a positive number when an answer
- The group also discussed this MetaSim error and the amount of additional information that will be needed to diagnosis it.