changes

一个很有用的EM算法程序包

03-Jul-07     * Corrected a problem in PEIGS with recent Matlab
                implementations, in which eigenvalues were not
                necessarily returned in descending order.   

15-Mar-07     * Changed screen output of effective number of variables
                (peff) in REGEM for TTLS option. REGEM was erroneously
                writing out an effective number of degrees of
                freedom, rather than an effective number of variables
                (=truncation parameter). However, this was of no
                consequence for the performance of the algorithm. 

		Note that this definition of the effective number of
                variables is not precise for TTLS. A better definition
		would be based on the filter factor formulation of
		Fierro et al. (1997), but obtaining these would
		involve additional computation, which seems
		unnecessary for this merely diagnostic output. 

01-Jul-02:    * Users reported problems due to changes in the calling 
                sequence of EIGS in a beta version of an upcoming Matlab 
                release. 

08-Feb-02:    * Minor changes in PEIGS and GCVRIDGE to improve downward 
                compatibility with Matlab 5. Switched to multiple ridge 
                regressions as default regularized regression method.

02-Jan-01     * Adaptation to Matlab 6 (minor changes in PEIGS). REGEM now 
                has an additional optional parameter OPTIONS.minvarfrac, 
                from which an upper bound on the regularization parameter
                can be constructed. 

04-Apr-00:    * CovRes in REGEM is now a full matrix, no longer a
                sparse matrix. Not allocating memory at initialization
                of the sparse matrix had significantly slowed down the
                algorithm.

22-Mar-00:    * All variables in a dataset are now scaled at the
                beginning of an EM iteration and regularization in
                standard form is performed. Before, in each individual 
                record only the variables with available values were 
                scaled. Scaling all variables slightly changes the
                objective of generalized cross-validation: instead of
                estimating the regularization parameter for which the
                expected *ms error* of the imputed values is minimum,
                GCV now estimates the regularization parameter for
                which the expected *relative ms error* of the imputed
                values is minimum. REGEM with MRIDGE might thus produce
                slightly different results; however, REGEM with IRIDGE
                should produce the same results as before.