Generic Statistical Data Types
It is clear that if we want different statistical
environments to communicate data between each other
that each must have an ability to understand the other's.
There are a variety of ways of organizing this.
We think that IDL is a suitable non-intrusive way
of declaring properties of a type without requiring
implementation or representation in that way.
The idea is that IDL declares what attributes
and methods are available to a client in order to retrieve
or modify data in a remote object, but says nothing about
how different systems must implement this.
This avoids the age old problem of agreeing on
language definitions of data types that are a
compromise between numerous systems and that are different
from all.
Additionally, since IDL is object oriented and
supports class/interface inheritance, etc.
those who need a greater array of methods/access
can sub-class any common definitions at the risk of not being
understood by others.
And finally, since CORBA objects are self describing
in the sense that one can determine the available IDL
information from the Interface Repository,
systems can dynamically convert a remote object
from one type to its own representation with little or no
code specific to that data type.
(This is the basis of the S approach to CORBA
which we are currently developing.)
More reasons why we might use IDL or ILU
are presented in a general comparison of
So, now is the time for us to decide what common
data types we require in the statistical domain
and what their characteristics should be.
We hope that this will be an endeavour that
we can all work on and that no one language
and users of a language would dominate.
The Approach
All languages provide a set of primitive types
such as double, integer, character/string, array, etc.
From these we develop higher level
constructs which contain these primitive types.
And then we recusively include higher level structures
inside other higher level structure definitions.
CORBA/IDL provides a good set of primitive types.
From this we define interfaces/classes for
vectors (of types integer, double, string, etc.)
This gives us the vector/list-like structures we are familiar
with in the common statistical languages and gives us the basis for Iterator classes.
Similarly, we require an array of arbitrary objects (much like the
list in Lisp and S).
Next we should probably turn creating interface definitions for
more statistical data types such as
matrices (perhaps different types such as sparse, etc.),
factors, data frames, etc. and perhaps even formulae, etc.
Duncan Temple Lang
<duncan@research.bell-labs.com>
Last modified: Fri May 22 08:06:57 EDT 1998