COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Bruce,
C00009 ENDMK
C⊗;
Bruce,

Here is the tail end of the section on RLL.  THe first few pages are
circled on a hardcopy which I have given Ed.  This other material exists on
our Altos; if you tell me (i) how and (ii) that you want it, I will try to
somehow transfer it over to you at Sumex.

Regards,
Doug



Directions for Future research on RLL

The RLL system must be developed  into a usable package, and  experimented
with.  Only through multiple usages will directions for future research be
revealed.  Several systems are  already planned for  (some layer of)  RLL,
including:  WHEEZE  (Smith  &  Clayton: diagnosis  of  pulminary  function
disorders), ROGET  (Bennett: guiding  a physician  in constructing  a  new
expert system automatically, by employing units for Diagnosis in  general,
for Rule Acquisition, etc.), and a few non-medical applications.


Already, we have isolated several core research issues, which will  govern
the direction of our research during the next five years.  This agenda  of
issues includes:

(1) Incorporating other  researchers' representational  schemes into  RLL.
For instance, the user should  be able to specify that  he or she wants  a
KRL-like environment,  or  a MYCIN-like  environment,  and the  bundle  of
"organ-stops" which must be adjusted should change immediately.

(2) Codifying knowledge about  representation. This includes refining  our
taxonomies of inheritance modes, control structures, etc.

(3) Building  up  our  stock of  ideas  about  fundamental  representation
issues:  dealing with nested quantification, mass nouns, time, intensional
objects, counterfactual conditionals, etc.

(4) Easier knowledge acquisition.  One approach to this is to improve  the
interface to  an expert  user,  who must  transfer  his knowledge  into  a
program.  For example, J. Bennett's program ROGET, mentioned above,  which
can direct  such a  knowledge  acquisition process  because it  possess  a
detailed model of what comprises such a session.  A second, and  currently
underexplored, approach is to have the program automatically discover  the
knowledge for itself.  This may appear  much more costly, but recall  that
"expert knowldge" breaks  down into  (i) facts and  (ii) heuristics.   The
latter are almost  never articulated by  experts; it is  easier to  induce
them from examples.  This leads us to study:

(5) Automatically discovering new  domain-dependent heuristics.  This  was
the critical lack in the earlier  AM system [ref], which had some  success
in  autoamtically  discovering  new   (albeit  elementary)  concepts,   by
combining old ones.   Our work in  the past two  years has indicated  that
powerful heuristics can be found as simple patterns in he values of slots,
provioed the system has very  useful domain-specific slots.  Thus this  is
pointing us to the problem which follows:

(6) Automatically  discovering  new  domain-dependant  slots  which  prove
useful.  E.g., after proving the fundamental theorem of arithmetic, decide
that PrimeFactors is a useful slot for any (unit representing a) number to
have.  Our  approach,  as usual,  is  to  explicate and  codify.   We  are
building a taxonomy of slots;  i.e., of useful relations between  concepts
(units). Already the number of slots is in the hundreds, and over the next
five years  we  expect this  number  of  different kinds  of  slots  worth
distinguishing to increase by an order of magnitude.  This in itself  will
raise several  new issues  to deal  with, which  were invisible  when  the
number of different slots was under a dozen.

(7) Ultimately,  tackling the  probelm  of automatically  discovering  new
representations of knowledge.   Currently, our  only plan  to attack  this
problem is  to represent  each type  of representation  (e.g.,  graphical,
schematized, linguistic,...) as a unit,  organize these into a  hierarchy,
and see if  the domain-independent  heuristics are adequate  to guide  the
search for new and better representation schemes.