taalbrecht / multieqoptimizer Goto Github PK

Ways to speed up determinant calculations:

Discrete choice:

1. Pass all candidate points to calculation at once and perform updates per the row being modified instead of recalculating the whole matrix each time

Consider adding optimizer search feature that increases d-efficiencies by slightly varying weights for each formula away from target weights to cause design to escape local minima. "Vibrate" the design out of a local optimal point. A few preliminary tests seem to indicate that this may, in fact, be beneficial. Also consider altering gibbs and federov sampling.

The optimizer will crash with the following error when a design matrix is encountered that produces a degenerate matrix. Somewhere in the determinant calculator for the choice model, this must produce an issue. Find error and catch to make sure it simply returns a determinant of zero instead to prevent the crash.

Error in if (obj_temp < obj_current) { :
missing value where TRUE/FALSE needed

This error is produced if a formula contains only one term:

Error in rep(0, ncol(CurrentMatrix) * ncol(CurrentMatrix)) :
invalid 'times' argument

Fix namespace loading of nlme. Currently the user must load nlme prior to running functions from this package

May want to randomize search order for federov and gibbs sampler since they now make movements if there is simply no loss in efficiency. This can result in the design getting "stuck" on the last few rows of the candidate set or on the edge of the design space.

algebraic_range function will not find appropriate min and max for contrast.sum coded factor combinations (finds 0 as minimum instead of -1).

Add factor compatibility code from Federov algorithm in Discrete choice optimizer code to gibbs sampler and linear d-optimal optimizer.

Change code to require labeling of prior vector effects with variable name to ensure that the proper prior is applied even if the R code reorders factors levels and/or variable orders when the model matrix is created