1.50-b1

	- Synched configs with 1.45-b1. Except no edge pixels are ignored
	under -deep (deep.cfg). Re-adjust as needed...

	- Slightly stricter temporal pixel analysis chi^2 activation condition
	than 1.45-a2.

	- Row medians erroneously returned a source correction factor...

        - [Cleaned up unneccessary robust estimation code -- inherited from
        1.50...]

	- Row and Block medians were removed from data with incorrect gains.
	Now fixed.
	
	- RowBiasDrifts has added robust estimator option, controlled as usual
	with -ErowDrift= option or ROW_DRIFT_ESTIMATOR key. 

[1.45-b2]

	- [Previous 1.45 releases still had MUX variance bug.]

	- [Cleaned up unneccessary robust estimation code -- inherited from
	1.50...]

[1.45-b1]
[1.44-4]
	- Beam size was not read back by ReducedData, defaulting to 8.5 arcsec
        size at all times.

	- Better reductions on a larger range of sources resulting from new
        stricter chi^2 requirements for despiking and temporal pixel analysis.
	
	- Crush -correct= (FLUX_CORRECTING_FWHM) still misbehaved if
        'none' was argumented. Now fixed.


[1.45-a1]

	- Added alternative model for row drifts. RowBiasDrifts can be used
        instead of RowPolynomials. The new model is based on the assumption
        that the drifts are caused by slight fluctuations in the bias voltages
        to rows. Accordingly, the gains by which these couple to pixel readout
        are different than what is assumed in the Polynomial model (which
        essentially assumes that row drifts are temperature fluctuations.).
        This model is controlled, in the usual way, by the keys ROW_DRIFT_TURN
        and ROW_DRIFT_T and the corresponding -IrowDrift= and -rowDriftT
        options.
	
[1.44-3]

	- The crush -correct= option did not work as desired. It was rather
        botched. FLUXES FROM PRIOR 1.44 RELEASES ARE NOT RELIABLE.

	- Corrected MUX variance calculation.

	 - MUXOffsets and now revise the model and will only keep model for rows
        which do appear to have detectable model levels. This is good because
        MUX offsets seem to come ang go and affect only a few rows at once...

	- Windows .bat batch files added for easier Windows usage. Simply
        unzip windows.zip, then run 'cleanup' (cleanup.bat). The UNIX scripts
        will be replaced with Windows batch files to the same effect!!!

	- Deconvolution now makes the replacement beam size as the image
        beam size, and recalculates the necessary constants.

1.50-a2
1.44-2
	- MUX RMS is estimated only for rows that have a detectable MUX
        term.

	- Changed pixel weighting frequency band to 0.7-0.9 times the 
	available range of frequencies. This way any high frequency trouble
	(like filtering or 13Hz noise will not bias pixel weights.

	- Console flux summary now in incident Jy/beam unit instead of detector
	nV. This way the numbers are more physically meaningful. Also shown
	the estimated residual source pollution (Jy/detector) by interfering
	models(!).	

	- More intelligent pipeline ordering based on estimated source
	pollution effects of various models. Does not use source diameter any 
	longer.

	- Map weights no longer displayed for scans that were did not
	contribute to the composite map. Instead '[--]' is displayed to
	indicate the scan was not used. Those scans are also assigned
	zero weights.		

	- A wrong normalization found in determining whether it is safe or not
	to solve for row polynomials and MUX offsets for extended sources. 
	This deteriorated the reductions of faint extended sources. Now 
	corrected. 

	- Relaxed flagging and map clipping directives for faint and deep
	reductions (faint.cfg & deep.cfg).

	- Deep reductions (deep.cfg) no longer filter large scale structures.
	However, such filtering will occur on the maps themselves until
	faithful source extraction is required (-Ifidel= flag). Normally,
	this will only affect the first few map generations but not the final
	output. Occasionally one may want to enable the RegionalCorrelations
	model to perform such filtering.

	- Faint and deep reductions (faint.cfg and deep.cfg) no longer
        exclude array edges. Though in some cases the user may want to
        re-enable the edge clipping (e.g. -edge=4:4:0:0)

	- -faint and -deep (faint.cfg, deep.cfg) now fit second order sky
	polynomials.

	- AccelerationResponse re-activated in pipeline for -faint and -deep
	reductions (faint.cfg, deep.cfg). Activation condition chi^2 
	requirement relaxed to < 2 for 'auto' activation.

	- Altered MUX Offset modelling. While the previous modelling was
        technically correct, it did not perform well on non-intact rows.
        The new estimator measures half row offsets separately first and
        estimates the MUX offset with the difference over the two halfs,
        rather than getting the MUX offset directly with a single estimator
        as done before. The result is expected to be more reliability for
        this model.

	- Improved error function integral estimation to use with significance
        estimation in 'detect' tool.


1.50-a1
	- Changed some default configuration settings (crush.cfg, faint.cfg
	and deep.cfg) from the previous version.

	- MAP_SMOOTHING key -smooth= option have new values. Smoothing
	can appear at three stages, internally in the reduction (i) while
	initial source map is constructed, (ii) internally after faithful
	map construction is set and (iii) for the final map. Accordingly,
	the directive can be 'initial', 'faithful' and 'final' to determine
	which stage is set. Additionally 'internal' sets the first two internal
	stages at once, while 'always' sets all stages. The value 'none' or
	'off' disables all smoothing. Thus
		
		-smooth=internal:6.0
	
	will smooth both maps at both internal stages by 6.0 arcsec FWHM. 
	
	- Added robust estimnators for PixelOffsets, CorrelatedNoise,
	PixelDrifts, RowPolynomials, MUXOffsets and BlockCorrelations. These
	alternative estimators are based on medians, rather than weighted
	means, and consequently will be less biased by the presence of compact
	source structures in the data steam. The estimator to use can be 
	selected via the XXX_ESTIMATOR keys or -Exxx= option flags, where
	XXX or xxx stand ofr the model ID's just like in the XXX_TURN and 
	-Ixxx= keys and flags. The valid values are 'maximum-likelihood' or
	'robust'. E.g. to use robust estimators instead of the default
	maximum-likelihood ones for correlated noise estimation, you can 
	use either:
		CORRELATED_ESTIMATOR = robust
	configuratiuon key, or
		crush [...] -Ecorr=robust [...]
	option flag.		
	Robust estimators on BlockCorrelations may be desirable. However,
	these did not seem fully reliable, and therefore the default 
	maximum likelihood estimators are used instead.

	- Old regional correlation model is brough back under the alias of
	BlockCorrelations. This is a much faster way of estimating 
	correlations than the FFT based RegionalCorrelations. It also has 
	the ability of using robust estimators instead of maximum-likelihood 
	ones.

	- Improved error function integral estimation to use with significance
	estimation in 'detect' tool.

	- Very slight error in calculating the correlated noise uncertainties
	Does not effect reduction in any significant way.

