COMMENT ⊗   VALID 00004 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002		scientific goals, application goals for AI
C00009 00003	HAND/EYE RELATED APPLICATIONS
C00014 00004		scientific goals of AI
C00021 ENDMK
C⊗;
	scientific goals, application goals for AI


One focus of applications of AI research is to software systems which
cannot now be  built because of lack of predictability, managability,
and economic feasibility.  

As in the story about the intelligent mule, ARPA has  our attention. 
We need some  information and a lot of feedback  about what DoD needs
are and  whether our suggestions are at all relevant.  AI researchers
seem ready to respond to DoD needs, and lack a clearly-defined set of
possibilities. 

		APPLICATIONS OF AI

AI  now  directs  its  major  efforts   at  programming  systems  for
applications.   Hand/eye  attacks  applications   in  automation  and
teleoperator control.   Problem-solving  areas address  more  general
programming.    The more  general  the  programming system  the  more
distant  the payoff.  From our perspective,  we see a great advantage
in making advances within  well-defined areas.  We suggest  a typical
application below, and solicit  feedback about whether it and related
problems meet DoD needs. 

WE PROPOSE A UNIFIED  EFFORT TO DIRECTLY ATTACK SPECIAL  AREAS OF DoD
SOFTWARE WITH  SPECIAL PURPOSE SYSTEMS WHICH  EMBODY THE KNOWLEDGE OF
SYMBOLIC MATHEMATICS,  NUMERICAL MATHEMATICS,  AND PHYSICS,  TOGETHER
WITH SIMPLE PATTERN MATCHING TECHNIQUES OF NATURAL LANGUAGE. 

The DoD has a  vast programming effort.   From the outside, we  guess
that  there   are  problems  which   DoD  cannot   undertake  because
programming of moderate-size systems is not predictable, controllable
and economical.  A successful incremental improvement  in programming
systems would  make possible things which  are not possible now.   We
maintain  that  while some  improvement  can be  made  in programming
systems  with   current  technology,  current   systems  are   almost
exclusively syntactic,  in the sense  that the semantics  are trivial
(their data structures are only real, integer, boolean, vector).   To
make significant  improvement in programming  ease, the  systems must
have semantics of the special domain of the  program.  At the moment,
structured programming is a  name in search of an idea.  That idea is
detailed  semantics  of  the   problem  domain.    Most  effort   has
concentrated on the program domain. 

A class of programming to consider is scientific programming, typical
of orbital and  control programming.   We assume  these programs  are
important  to  the  military,  and  we see  that  the  AI  technology
associated   with   MATHLAB   and  MACSYMA,   STUDENT,   ie  symbolic
mathematics,  coupled with  numerical  mathemtics  are  now  adequate
foundations to build an incremental improvement to  FORTRAN, etc.  In
this  domain, physics defines the  semantics, and mathematics defines
the representations. 

A device operation is a  sequence of maps from one semantic structure
to the  next.   A program  is a  sequence of  maps from  one internal
representation to the next.  We must represent the semantic structure
in some internal  representation to do effective computation.  If the
internal representation is effective the program can be  very simple.
That is, the science is usually simple when expressed at the level of
the  problem  domain.   Now,  however, at  the level  of  the program
domain, the program may  spend pages setting up the  data structure. 
To simplify  the process of programming, the  most effective approach
is to  model the  data structures  and the  maps for  each  important
application domain.  These representations now have to be coded fresh
for each problem, in non-standard ways.  
	PROBLEM DOMAIN			PROGRAM DOMAIN
	semantic structure 1		data structure 1
		|				|
	      map 1			      map 1
		|				|
		↓				↓
	semantic structure 2		data structure 2
		|				|
	      map 1			      map 1
		|				|
		↓				↓
	semantic structure 3		data structure 3


More  generally,  we  see  this  approach  as  MORE-NATURAL  LANGUAGE
approach  to  programming  systems.    Wherever  a  compact  body  of
programming exists,  the semantics  can  be encoded  in this  way  to
simplify  programming.    The  ability  to  represent  other  domains
increases,  in a cascading fashion; each  new domain can use elements
from the others.  
HAND/EYE RELATED APPLICATIONS

  1) COMPUTER-ASSISTED TELEOPERATORS FOR HANDLING DANGEROUS MATERIALS
  I  have in  mind ordnance,  chemical,  biological, and  radioactive
  materials.   What is the advantage  of computer interactive systems
  for handling dangerous  materials?  First, let  me make clear  what
  sort  of  system  I  have  in  mind.    This  is  a  computer-aided
  teleoperator system,  with some stand-alone ability.  The system is
  an interactive system which augments teleoperator capabilities.  It
  is useful  for repetitive operations, and for  operations for which
  operator error in  repetitive would have  messy consequences.   The
  computer   would  give   the   advantage  over   teleoperator   and
  "teach-mode" control in:

    a) Computers can connect directly with sensors.   It is difficult
    to interface humans with sensors in an effective way. 

    b) Transformation  from  a  convenient input to  arm geometry for
    small arms, large arms, arbitrary geometries

    c) Repetitive execution saves  operator effort and cuts  mistakes
    caused by operator errors. 

    d) Smoothing and improvement of trajectories by computer. 

    e) Interactive editing of task sequences

    f)  Increased reliability  from  self-calibration and  continuous
    update of calibration. 



  2) AIRFRAME ASSEMBLY
  3) PARTS PROGRAMMING FOR AIRFRAMES
    There  are systems  under  development  for automation  of  parts
    programming  for NC  machining of  simple  shapes.   Our advanced
    facilities for representation  and graphics  make it possible  to
    automate  machining   of  complex  parts,  typical   of  airframe
    structural members. 
  4) DRAWING AND LAYOUT FOR MECHANICAL PARTS
  5) HARDWARE VERIFICATION
    Most DoD procurement  involves small volume  electronic systems. 
    We have  underway a project to verify  the functioning of digital
    hardware.   We have  a  record of  accomplishment in  this  area.
    Digital Equipment  corporation is  using our drawing  program for
    production. 
	scientific goals of AI



From my  perspective,  AI is  that domain  which  considers both  the
PROBLEM DOMAIN  and the PROGRAM DOMAIN.   Our scientific goals are to
solve real world problems which are not trivial.  That is,  a program
is  a series  of maps  between internal  representations.   Thus, our
concern  is to find  semantic representations of  problem domains and
find   equivalent   computer   representations   of    the   semantic
representations and the maps between semantic representations.