COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Cordell,
C00007 00003	∂26-OCT-74  0717		network site ISI
C00021 ENDMK
C⊗;
Cordell,
John mentioned your exchange with Lick and suggested another one,
which follows.

∂25-OCT-74  2326		S,LES
 Dear Lick:
 
 I heard third-hand that you are collecting bullets again.  Here is
 a list from our latest annual report (in press).
 
 
 		STANFORD ARTIFICIAL INTELLIGENCE PROJECT
 
 The work of the Artificial Intelligence Project has been basic and
 applied research in artificial intelligence and related fields, such
 as mathematical theory of computation.  Here is a short list of what
 we consider to have been our main accomplishments during the past year.
 
 			Robotics
 We have developed a two-arm synchronized manipulation capability and
 tested it on several mechanical assembly tasks that are beyond the
 capability of a single arm.  A new high-level "hand language" called
 HAL has been developed for specifying advanced manipulation tasks.
 
 			Computer Vision
 We have used near and far field stereo vision and motion parallax to
 locate objects spatially and to automatically generate contour maps.
 Another program can recognize things of the complexity of a doll or a
 hammer in various positions, using a laser triangulation system.
 
 		Mathematical Theory of Computation
 Using our LCF proof-checker, we have produced an axiomatization of
 the programming language PASCAL.  This represents a major step toward
 using LCF as a practical program verification system.
 
 			Theorem Proving
 An interactive system has been developed for structured top-down
 programming in PASCAL.  It guides the user in constructing a program
 in successive refinements and in proving its correctness.
 
 			Automatic Programming
 A successful new automatic programming system accepts descriptions of
 library routines, programming methods, and program specifications in
 a high level semantic definition language.  It returns programs
 written in a subset of ALGOL that satisfy the given specifications.
 Experimental applications include computing arithmetical functions
 and planning robot strategies.
 
 Another system works with algorithms expressed in a higher-level
 language and automatically chooses an efficient representation for
 the data structure.  It then produces a program that uses this
 representation.  Representations considered include certain kinds of
 linked-lists, binary trees, and hash tables.
 
 		Natural Language Understanding
 A system called MARGIE was completed that links natural language
 understanding, inference, and generation of natural language output.
 This is the first such system with a strong theoretical basis, namely
 "conceptual dependency".
 
 			Training
 During the year, six members of our staff published Ph.D. dissertations
 and another 32 graduate students received direct support.
 
 CC: LICKLIDER%ISI

∂26-OCT-74  0717		network site ISI
 Date: 26 OCT 1974 0717-PDT
 From: LICKLIDER at USC-ISI
 Subject: Bullets
 To:   LES at SU-AI
 cc:   Licklider
 
 	Thanks very much for the bullets.  What I had in mind when I
 called was mainly automatic programming, but I am very happy to have the
 broader coverage.
 
 	Here is some reaction/response, intended to contribute to the
 shaping up of this bullet business, which plays (or should/could play)
 an important role in keeping ARPA Hq. informed of and happy with what
 IPTO contractors are doing:
 
 Robotics
 
 	What, specifically, is the two-arm system capable of doing?
 Examples might be sufficient.  A general specification of capa capability
 illustrated with examples would be great.  Our readership delights in
 specificity.
 
 	How much does stereo improve performace?  How does far-field
 stereo vision work?  (In human vision, I thought stereo worked only in
 near field.)
 
 MT of C
 
 	Could you lay out the steps to a practical program-verification
 system, with projected times of achievement, so we can see where
 the LCF-PASCAL program stands in the over-all program?
 
 [Sorry I missed a heading.  'Computer Vision' should have preceded
 'How much'.]
 
 Theorem Proving
 
 	Gratified to see we have in over-all IPTO program at least
 one effort to create computer-aided top-down programming aid.
 It will be important to make measurements to see whether/how much
 this beats 'ordinary' programming.  Could you be sure the report
 gives a thorough treatment of this one?
 
 AP
 
 	From Cordell, I have descriptions of three AP programs, of
 which the second you mention is one, but I do not recognize the first
 you mention.  I am now trying to perfect a set of AP bullets.  Would
 you augment the description of this first one.  I'll send my
 descriptions of the other three for improvement by you and 
 Cordell.
 
 NLU
 
 	This is too general for our use but sounds as though a
 more specific description would be very useful.  Best supress the
 name.  Please explain conceptual dependency.
 
 Training
 
 	We are working (slowly) on the problem of the capacity of
 the field of AI to use research support effectively.  Ideally, we
 would have a model of the field that would include numbers of people
 at each of several levels of capability and cost, rate of movement
 from one category to another (e.g., rate of production of Ph.D.s
 and rate of dropping out of graduate study and dropping out of field
 altogether), etc.  Let me ask you to think about this matter a bit
 and indicate what you can tell us that will be helpful.  Our motivation
 is simply that we are planning to put more emphasis on and money
 into AI work, and the question is asked, 'Where are the competent
 people going to come from to do the work?'.
 
 	
 	Please let me know of any questions that my comments raise
 in you mind.  I want to develop, over the months, a way of getting
 good, informative bullets with the least pain to all concerned.
 Perhaps the best general guidance I can give is that the intended
 readers insist on understanding, are not satisfied with someting
 that just sounds important.  The problem for me is that (1)
 the items have to be pretty short but (2) have to carry enough
 context that a gerneralist in technology can understand them and
 their relation to DoD problems.
 
 				Regards
 
 				Lick
 
 P.S.:  I forgot to carbon-copy Cordell.  Would you please show
 this to him?  Lick
 -------