This file lists the major changes of BACK V5.* with respect to
previous version of Release V5.

OVERVIEW

 o  Improvements wrt V5.1
	- port to SWI-Prolog
	- improved abstraction of atmost(0,R) terms
	- improved treatment of features
 o  Improvements wrt V5.0
	- further unification of retrieval/single pattern 
	  <projection/tuple-generator> FOR <args>
	- describe/ describe_fully
	- dedicated tboxget (upward-compatible with V4)
	- full support of TBox/IBox/ABox revision
	- considerable speed-up
 o  Language Changes
 	- changed semantics of i-links
	- abandoned 'anyrole'
        - abandoned 'inherent_feature'
 o  Non-deterministic failures - SOLVED
 o  Syntax Updates
 o  Missing
	- getall <role>
	- complex role-terms in ABox expressions
	- persistence of query concs
	- difference


IMPROVEMENTS OVER BACK V5.1

 o 	BACK V5 is now ported to SWI-Prolog. SWI-Prolog is a public 
domain version implementation of Prolog developed by Jan Wielemaker,
University of Amsterdam (email-contact: jan@swi.psy.uva.nl). It is
available via ftp@swi.psy.uva.nl. Consult file Readme.Back52 for 
installation. 


 o	 The abstraction of atmost(0,R) terms has been made more
complete than in V5.1. An atmost(0,R) term is now abstracted for an
object x if none of the potential fillers for R can consistently
become an instance of the range of R or x's value restriction of R.
The implementation is not complete though; there exist cases in
which an R-filler would become inconsistent with the range or VR,
but BACK does not detect this and thus cannot abstract atmost(0,R).
Also, we haven't yet extended the solution to cope with arbitrary
atmost(n,R) terms.


 o	Feature processing has been improved: Features are classified
more correctly, and Subroles of features and compositions of features
are detected to be features too.  Features are now introduced as quasi
primitive subroles of the specified roleterm (if this is not a feature
itself).  Note that features defined as 
	"f := r and domain(atmost(1,r))" 
are not detected as features.


 o The name and the location of the dump file that's automatically created 
during tbox revision can be set via the variables 'tboxdumpfile' and
'tboxdumpdir' using the backstate/1 command.


 o	There have been minor fixes of errors reported by BACK users.



IMPROVEMENTS OVER BACK V5.0

  o	The retrieval interface has been further unified.
There is now a uniform interface to retrieve most of the things that
you might be interested in:

	backretrieve( VAR = <projection/tuple-generator> FOR <args> ).

The 'VAR =' is optional; without it the result of the retrieval will
be pretty-printed to the current output stream.
The <projection/tuple-generator> is a (Prolog-) list that specifies
which information should go into an output tuple for every of the
arguments in <args>. We had also called these "output functions"
or "actions".
<projection/tuple-generator> contains all the known things like
self, msc, rf/1. Also defined_as, describe and describe_fully now 
have to be put into the tuple specification. As new features we also 
provide nr(<Role>) and vr(<Role>) to ask for NRs and VRs.

Arguments (<args>) are either given explicitly by enumerating
them in a Prolog-list, or are objects retrieved by a getall query.
Examples are:

	backretrieve([defined_as, describe] for [father, has_son, peter]).
	backretrieve([self, msc, rf(r), vr(r)] for getall concept23).

Not all output functions can be applied to all entities. E.g., the msc
function is not defined for roles. backretrieve will inform 
you by an appropriate message in the result.

Some of the output functions can be applied directly on arguments.
These are all those which are atomic/have no argument:
describe, describe_fully, defined_as, introduced_as, self, msc.
Example
	backretrieve(describe paul).
	backretrieve(defined_as [father, has_son, peter]).


  o	describe and describe_fully are supported by V5.1.  In general,
descriptions obtained by describe should be minimized compared to
descriptions obtained by describe_fully; this is the case for concepts
and objects.

For attributes and numbers, describe and describe_fully yield the same
results; this is intendend. 

At the moment, the implementation of describe for roles does not
minimize every description as good as it should (which we hope to be
able to improve in future releases).  As a sideeffect of the weak
implementation of describe for roles, some roles in concept/object
descriptions will not be unfolded to the original names. Instead,
internal role objects will be returned, a description of which then
may be obtained via backretrieve.  However, if a name has been
introduced for a role or its inverse, either the role's name or inv(.)
will be used.


  o	There is (again) a dedicated programming interface (PIF)
accessible via the tboxget predicate. It is upward-compatible with V4,
has been extended for i-links, and should allow for access to most
information of interest that is not accessible via backretrieve or backask.
In the future we may merge tboxget and backretrieve to a single uniform
interface. 


  o	Revision/changeability of a BACK knowledge base is now fully supported.
This means that you can change any part of the knowledge base (ABox) and
knowledge base schema (TBox/IBox) at any time, and that you can enter
information in a mostly arbitrary order (e.g. you can define new concepts,
roles or i-links after you already have inserted objects).

Two remarks are necessary here: Changing objects (by forget or redescribe)
is supposed to be nearly as efficient as telling the corresponding information.
Its implementation should be stable.
Changing the schema (concepts, roles, i-links) after objects have been
introduced is NOT SUPPOSED to be EFFICIENT. It should be faster than
doing a backinit and reading everything anew, but it should be used
with caution. Remember it is a schema change. Thus we recommend that you
stay with the standard order TBox-IBox-ABox whereever you can.
Introducing a new concept is less expensive than introducing a new role,
and changing the definition of an existing concept/role is supposed
to be worst.


  o	V5.1 is significantly (!) faster than its predecessor; if you
had installed V5.0.* you will notice immediately. Due to the higher
expressivity, however, the overall performance of V5.1 is still a bit
slower than V4.


LANGUAGE CHANGES

  o	We have weakened the semantics of i-links: I-links
are not applied any longer to value-restrictions. Since VRs are
forward-propagated to fillers, the i-links will be applied on
the fillers, though (see below).
The new behavior corresponds to the treatment of i-links in CLASSIC.
Examples:
	c1 :< anthing.
	c2 := all(r,c1).
	c3 := c1 and atleast(2,s).
	
	c1 => c4.
	
	c1 ?< c4.   		%% yes
	c3 ?< c4.   		%% yes
	\+ c2 ?< all(r,c4).  	%% no: 'c1 => c4' is not applied on the VR

	x1 :: c3.
	x2 :: c2 and r:x3.
	x4 :: c2 and atmost(1,r) and r:x5.
	x6 :: c2.

	x1 ?: c4.		%% yes
	x3 ?: c4.		%% yes due to forward prop. of VR from x2.
	x5 ?: c4.		%% yes due to forward prop. of VR from x4
	\+ x2 ?: c4.		%% no: filler set not closed
	x4 ?: c4.		%% yes: filler set closed
	x6 ?: c4.  		%% no: 'c1 => c4' is not applied on the VR
	

  o	The top-most role 'anyrole' has been abondoned. The reason is
that 'anyrole' allows semantically much more inferences than we
planned to handle or could handle efficiently (e.g. it would have been
possible to express that there could not be more than n instances of a
concept).  We opted for not making BACK even more incomplete.

Primitive roles now are introduced by giving either their domain and/or
range explicitly. If a role is introduced as 'r:<domain(anything).'
the default range(anything) will be enforced for r.


  o	The type 'inherent_feature'  has been abondoned. Inherent
features were meant to be features which always have exactly one role
filler, in contrast to the type 'feature', which has at most one
filler. The reason is basically that we have found it difficult to
handle correctly the implicit disjunction introduced by inherent
features. 


NON-DETERMINISTIC FAILURES - SOLVED

Since Version 5 we observed that in large BACK knowledge bases
or in our 'alltests' test-suite BACK failed in unexpected situations,
and that we could not deterministically reproduce these situations.

This problem was due to an error in the way Quintus
Prolog Version 3.1 was dealing with memory allocation and/or garbage
collection. The problem has been solved with Quintus Prolog 3.1.3.



SYNTAX UPDATES

There have been minor changes of the syntax of 5.1 compared to 5.0.*.
Some of them have been mentioned above. In any case, consult the
updated version of the file 'Doc.Back51'.



STILL MISSING

  o	A simple operator for retrieving the extension of a role.
In BACK V4 there existed something like 'getallrel'. First we planned
to allow for 'getall <role>' but this did not fit well into the
unified backretrieve format. If you are interested in getting the 
extension of a role 'r' ask

	backretrieve([self,rf(r)] for getall anything).

You could improve the query (narrowing down the result set to 
object that actually have a r-filler) if you provide domain  and range
of the role:
	
	backretrieve([self,rf(r)] for 
			getall <r-domain> 
			and r:someknown(<d-range>)).


  o	The possibility to use arbitrary role terms in ABox tells and
queries. The use inv(<role>) is allowed, though, if the range of the
role is subsumed by anything (as in V4).

  o	Optional persistency of query concepts.

  o	Difference as a backretrieve operator.