perm filename AI72.QUA[ESS,JMC]1 blob sn#005567 filedate 1972-04-05 generic text, type T, neo UTF8
00000	
00100	
00200	
00300	
00400	
00500	
00600	
00700	
00800	
00900	                         STANFORD UNIVERSITY
01000	
01100	
01200	                     COMPUTER SCIENCE DEPARTMENT
01300	
01400	
01500	                            April 1, 1972

01600	
01700	
01800	
01900	
02000	
02100	
02200	
02300	
02400	
02500	                    Ph.D. QUALIFYING EXAMINATION

02600	
02700	
02800	                       Artificial Intelligence

02900	
03000	
03100	
03200	
03300	
03400	
03500	
03600	
03700	The examination will be open book.  The first session will be from
03800	9:30 to 12:30 pm, and the second session will be from 1:30 to 4:30
03900	pm.  No work on the examination will be done during the lunch break.
04000	
     

00100		1. a.  In terms of the quality of solutions  found,  and  the
00200	efficiency  with  which solutions are produced, the Heuristic DENDRAL
00300	program is one of the most powerful heuristic programs in  existence.
00400	What is the primary source of this problem-solving power?
00500	
00600		b.   In  his  paper on Heuristic Programming, subtitled "Ill-
00700	Structured Problems",  Newell  introduces  concepts  and  terminology
00800	intended to categorize and describe heuristic programs.  Use Newell's
00900	concepts and terminology to describe Heuristic DENDRAL.
01000	
01100		c.  What are the purposes of having  a  systematic  generator
01200	(the DENDRAL algorithm) at the heart of Heuristic DENDRAL?
01300	
01400		d.  We use heuristic processes to achieve search reduction in
01500	administering the search for a solution to a problem.  How  does  the
01600	heuristic   process   known  as  the  Planner  in  Heuristic  DENDRAL
01700	contribute its heuristic power to search  reduction?   Illustrate  by
01800	making  reference to some of the results in the results tables of the
01900	DENDRAL paper you were asked to read.  From a heuristic search  point
02000	of  view,  how  does  "planning" in DENDRAL differ from "planning" as
02100	this method has been discussed elsewhere in the A.I. literature (e.g.
02200	the Planning Method of GPS, Hewitt's Planner, Robot Planning)?
02300	
02400		e.   You  were  asked  to  read a paper by Amarel in which he
02500	disucsses representation of knowledge and  shift  of  representation.
02600	How  has  this  problem  been  studied  in  the  context  of the task
02700	environment of Heuristic DENDRAL?  What are the results?
     

00100		2.   The unbounded unit-preference strategy is: "compute the
00200	resolvents of all unit clauses with every clause before computing the
00300	resolvents of any pair of non-units".
00400	
00500	The  input  clause  strategy is: "compute the resolvents of a pair of
00600	clauses only if one of them is a member of the initial set of clauses
00700	(i.e. an axiom or the negation of the theorem)".
00800	
00900		a.   Give  examples showing that both of these strategies are
01000	logically incomplete.
01100	
01200	
01300		A  replacement  rule of inference for equality may be defined
01400	as follows.
01500	
01600		Let E be the equality predicate and s, t, u be terms.  Let  A
01700	and  B  be clauses with no variables in common such that A contains a
01800	positive equality atom, either A = E(s,t)∨A' or  A=E(t,s)∨A',  and  a
01900	term  u occurs at least once in B.  (Note: u may occur as a subterm).
02000	Let s and u have a common substitution instance, and suppose that α a
02100	most  general  unifier  such  that  sα=uα.   Let  B* be the result of
02200	replacing an occurrence of uα in Bα by tα.    Let  C  be  the  clause
02300	A'α∨B*.  Then C may be inferred by replacement from A into B.  Denote
02400	the set of such inferences by P(A,B).
02500		If A and B have common variables, these must be eliminated by
02600	a change of variables before the rule is applied.
02700	
02800	
02900		b.  For all A and B, is P(A,B)=P(B,A)?  Prove your answer.
03000	
03100	
03200		c.  Let A be E(f(x,g(x)),e) and B be  E(f(x,x),e).   Compute
03300	P(A,B) and P(B,A).
03400	
03500	
03600	
03700		d.  PROVE:  For any C that can be inferred by replacement
03800	from A and B there is a C' satisfying (i) C' implies C, and (ii)
03900	C' is obtained by a sequence of resolutions from the set consisting
04000	of A, B, and the axioms for equality.
04100	
04200	
04300	[section (d) of this question was omitted from the actual examination
04400	on April 1, 1972]
     

00100		3.  After reading the speech report for inspiration, you have
00200	accepted  a consulting job with the linguistics department to predict
00300	the feasibility of a speech understanding system.  You are given  the
00400	following  vocabulary  and grammar.  You want a computer to recognize
00500	semantically and syntactically legal sentences.  The  department  did
00600	not specify the semantics but any reasonable assumptions will do.
00700	
00800	The vocabulary is:
00900	
01000		programs, monkeys, termites,
01100		search, climb, eat
01200		trees, bits, bananas
01300	
01400	The grammar is:
01500	
01600		S → SUBJECT VERB OBJECT
01700		SUBJECT → programs, monkeys, termites
01800		VERB → search, climb, eat
01900		OBJECT→ trees, bits, bananas
02000	
02100	a.	First, assume a probability of correct recognition of one  of
02200	the  vocabulary  words, when isolated, to be .7.  Suppose the lexical
02300	segmentation scheme is perfect.  Without  use  of  the  grammar  what
02400	correct   string  recognition  rate  (all  words  correct)  might  be
02500	expected on three-word strings?
02600	
02700	b.	Based on the results of  the  speech  report  how  might  the
02800	probability of word-confusion error depend on vocabulary size?
02900	
03000	c.	Make  a  reasonable  assumption (either your answer to (2) or
03100	some other guess) of the effect of  vocabulary  size  on  recognition
03200	rate.   State  your assumption.  Now, using the grammar, but still no
03300	semantics, what correct string recognition rate  might  be  expected?
03400	(rough calculations are fine).
03500	
03600	d.	Specify  your  assumed semantically meaningful strings.  Show
03700	precisely why the recognition rate  is  better.    For  extra  credit
03800	calculate  an  expected  correct 3-word string recognition rate using
03900	both syntax and semantics.
04000	
04100	
     

00100		4.  Consider the following variant  of  the  missionary   and
00200	cannibals problem:
00300	
00400		"Three  missionaries and three cannibals come to a river that
00500	they wish to cross.  They find a boat that holds two people  and  can
00600	be  rowed  by one or two.  However, if one person rows by himself, he
00700	will be too tired to row by himself again.    Besides  that,  if  the
00800	cannibals  ever  outnumber  the  missionaries  on  either bank of the
00900	river, the missionaries will be eaten.  How can they all safely cross
01000	the river?"
01100	
01200		a. Write a LISP program to find a solution.
01300	
01400		b. Write a micro-PLANNER program to find a solution.
01500	
01600		c.  Write  a  situation-calculus description of the situation
01700	and the effects of actions from which it  follows  that  there  is  a
01800	solution.    The   "result"   formalism  of  McCarthy  and  Hayes  is
01900	recommended.
02000	
02100		d. Discuss the problem of making a program that could go from
02200	the  above  English statement of the missionary and cannibals problem
02300	to a LISP program for doing  the  tree  search.   Would  the  PLANNER
02400	formalism  or  the  McCarthy  and  Hayes  formalism  be  suitable  as
02500	intermediate steps.  Why or why not?  In so far as  possible,  divide
02600	the   overall   problem   into  sub-problems  that  might  be  solved
02700	independently.
02800	
     

00100		5.  One of the central problems in the recognition of  scenes
00200	involving  plane-bounded  objects is the segmentation problem.  Falk,
00300	in his thesis, suggests improvements to Guzman's algorithm.  Describe
00400	Guzman's  and  Falk's  algrorithms.   Give  an example different from
00500	those in Falk's thesis where  Guzman's  fails  and  Falk's  succeeds.
00600	Give an example where they both fail.
00700	
00800	Extra credit: 
00900		a. Extend Falk's algorithm to cover the  case  you  presented
01000	above.   If the new algorithm doesn' cover all cases, find a counter-
01100	example.
01200	
01300		b. Discuss the segmentation problem for curved objects.