This file is intended for further news about bug fixes and further modifications of the package.
The (initial) current version was 0.16-5.

Changes with 0.16-8
-------------------

1) Fixed numerical issues when calculating REML function values for badly conditioned covariance matrices

2) Added bootstrap option for estimating local nugget variances of Cholesky decomposed terms
   during the kriging approximation of variance matrices of statistics

3) Solving linear equations/doing matrix inversions is more robust by `gsiSolve` and `gsiInv` trying a Cholesky decomposition
   first and, if it fails, SVD or finally a Bunch-Kaufman factorization. This may lead to computations of generalized inverse
   matrices or only approximate solutions to the equations.

4) New rigorous checks of possible `NaN`, `NA` including non finite values added

Changes with 0.17
-------------------

1) testing an approximate root is done only in case of no root and if quasi-deviance is significantly greater than zero by a
   user-defined tolerance, i.e. `ftol_abs`.

2) C function for FD approximation now uses pre-allocated storage

3) Improved LS in quasi-scoring, now using norm of QS and QD as monitor functions (switching when no progress detected);
   no valid step signals a minimum step length event thus terminates with convergence code and not an error anymore.

4) If local search was successful, then (even during global phase) use it as the current iteration point

5) Start local search from last sample point (if at global phase) or last local minimum (if at local phase)
  unless multistart option is used, see below.

6) Use a last local search at maximum iterations if at global phase with highest weight 	

7) Added multi start support if the algorithm is in global phase or local phase. The best root found so far (for the criteria see the vigntte)
   is selected and used for minitoring the progress, e.g. whether to switch to the local phase again or termination. 
   COMMENT: Consistency checks and best root selection are now done only for "real" roots, e.i. quasi-score vector/quasi-deviance
   when these are zero in finite precision (acc. to given tolerances)	  

8) Added scaling option of parameter vector and quasi-score vector used in line search of quasi-scoring to test stopping conditions 		  
   
9) Function `qleTest` now additionally can be used to test a different parameter than the estimated one (including another observed
   vector of statistics). This also implements a multistart optimization.
		  
Current known issues
--------------------
	o flip of sign in quasi-score vector (qscoring)	  
		The implementation of the quasi-scoring iteration has some trouble with the correct
		sign of the quasi-score components for values near zero mostly due to finite precision problems
		
		
	
TODO
----	
	o provide 'nlm' routine: (Quasi-)Newton for Mahalanobis criterion function with constant variance matrix,
	 	gradient as quasi-score and quasi-information as Hessian    
	