COMMENT ⊗   VALID 00021 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00004 00002	%slides.tex[e84,jmc]	slides for Presidential Address
C00005 00003	\centerline{\bf What is Common Sense?}
C00006 00004	\centerline{\bf Feigenbaum's question:}
C00007 00005	\centerline{\bf Effects of actions in achieving goals}
C00008 00006	\centerline{\bf Moving and painting blocks using circumscription}
C00009 00007	$$\eqalign{∀x l s.¬clear(top(x),s) &∨ ¬clear(l,s) ∨ heavy(x) ∨ l = top(x)\cr
C00010 00008	\centerline{\bf Programs with common sense -- 1958}
C00011 00009	\centerline{\bf Theoretical knowledge}
C00012 00010	\centerline{\bf Meta-facts about appearance and reality:}
C00013 00011	\centerline{\bf Philosophical presuppositions of AI:}
C00014 00012	\centerline{\bf The principle of rationality:}
C00015 00013	\centerline{\bf Example of unnecesary elaboration:}
C00016 00014	\centerline{\bf The epistemology of common sense:}
C00017 00015	\centerline{\bf Areas of common sense knowledge:}
C00018 00016	\centerline{\bf Open questions about common sense:}
C00019 00017	\centerline{\bf The Advice Taker (1958):}
C00020 00018	\centerline {\bf Formalized non-monotonic reasoning (1978):}
C00021 00019	\centerline {\bf Formula Circumscription:}
C00023 00020	\centerline{\bf Common sense data base:}
C00024 00021	\end
C00026 ENDMK
C⊗;
%slides.tex[e84,jmc]	slides for Presidential Address
\magnification=\magstep3
\hsize=6truein\vsize=4.5truein
\openup 6pt
\nopagenumbers
\raggedright\pretolerance=10000
\def\itemb{\item{$\bullet$}}\def\itemx{\par\hang\indent\ignorespaces}
\def\itemxx{\par\indent\hangindent2\parindent\indent\ignorespaces}
\centerline{\bf What is Common Sense?}
\bigskip
A certain combination of knowledge and reasoning ability required
for successful behavior in complex environments.
\bigskip
\noindent I will discuss three things:
\medskip
\itemb what kind of knowledge?
\itemb examples of common sense knowledge.
\itemb what reasoning abilities?
\vfill\eject
\centerline{\bf Feigenbaum's question:}
\bigskip
Is common sense just a matter of
a very large number of pattern-action rules?
\bigskip
\centerline{\bf My answer:}

No!  Unless the expert system interprets a more
sophisticated knowledge using system.  If you do that you still have to
define the more sophisticated system and write your knowledge in its
formalism.
\vfill\eject
\centerline{\bf Effects of actions in achieving goals}

\noindent The subject of the most axiomatic work.
\bigskip
\itemb situation calculus.
$s' = result(e,s)$ is the situation that results
when event {\it e} occurs in situation {\it s}.

\itemb frame problem.
How to avoid specifying everything that doesn't change?

\itemb qualification problem.
How to avoid endless qualifications in axioms?
\vfill\eject
\centerline{\bf Moving and painting blocks using circumscription}
\overfullrule=0pt
$$∀x e s.¬ab(aspect1(x,e,s)) ⊃ loc(x,result(e,s)) = loc(x,s)$$
$$∀x e s.¬ab(aspect2(x,e,s)) ⊃ color(x,result(e,s)) = color(x,s)$$
$$∀x l s.ab(aspect1(x,move(x,l),s))$$
$$∀x l s.¬ab(aspect3(x,l,s)) ⊃ loc(x,result(move(x,l),s)) = l$$
$$∀x c s.ab(aspect2(x,paint(x,c),s))$$
$$∀x c s.¬ab(aspect4(x,c,s)) ⊃ color(x,result(paint(x,c),s)) = c$$
\vfill\eject
$$\eqalign{∀x l s.¬clear(top(x),s) &∨ ¬clear(l,s) ∨ heavy(x) ∨ l = top(x)\cr
&⊃ ab(aspect3(x,move(x,l),s)).}$$
\vskip .15truein
$$∀l s.clear(l,s) ≡ ¬∃x.(¬trivial(x) ∧ location(x,s) = l).$$
\vskip .15truein
$$∀x.¬ab(aspect7(x)) ⊃ ¬trivial(x).$$
\vfill\eject
\centerline{\bf Programs with common sense -- 1958}
\vskip .15truein
\noindent Represent as sentences of mathematical logic:
\itemb general facts about the effects of events
\itemb goals to be achieved
\itemb One should do what achieves one's goals.
\itemb facts about the particular situation
\vskip .15truein
\noindent Deduce $should(action)$ and then do {\it action}.
\vskip 0pt
\noindent Deduction must be supplemented by non-monotonic reasoning.
\vfill\eject
\centerline{\bf Theoretical knowledge}
\vskip .15truein
\noindent ``A container is sterile if all the bacteria in it are dead''.
\vskip 0pt
The knowledge is theoretical, because we don't use it directly by
\itemb checking each bacterium to see if it is dead
\itemb or knocking each bacterium on the head to kill it.
\vskip 0pt
\noindent Instead we put the contents in agar to test it or heat it
to sterilize it.
\vfill\eject
\centerline{\bf Meta-facts about appearance and reality:}
\vskip .15truein
\itemb We must infer from appearance to reality.
\itemb We jump to conclusions.
\itemb Our stable knowledge goes from reality to appearance.
\vskip .15truein
\centerline{\bf Facts about appearance and reality}
\itemb Dogs sometimes look for food in trash cans.
\itemb An overturned trash can was overturned by a dog.
\itemb Mischievous children sometimes overturn trash cans.
\itemb Fleeing burglars sometimes run into trash cans.
\vfill\eject
\centerline{\bf Philosophical presuppositions of AI:}
\vskip .15truein
In 1958 the linguistic philosopher Bar-Hillel remarked that
formalizing common sense involved philosophical presuppositions.
He was right.  For example, axioms about the relation between
appearance and reality assume that both exist.
\vfill\eject
\centerline{\bf The principle of rationality:}

\itemb It will do what it believes will achieve its goals.

\centerline{\bf An elaboration:}

\itemb It intends to do what it believes will achieve its goals.

\centerline{\bf Non-monotonic contractions:}

\itemb It will do what will achieve its goals.

\itemb It will achieve its goals.

\vfill\eject
\centerline{\bf Example of unnecesary elaboration:}
\vskip 0pt
``He believes that stepping on the brake will stop his car and avoid
hitting the car in front, so he intends
to step on the brake''
\bigskip
\centerline{\bf A non-monotonic contraction:}
``The car in front has stopped, so he'll hit the brake''.
\vfill\eject
\centerline{\bf The epistemology of common sense:}
\bigskip
\itemb What must a robot know about the common sense world?

\itemb What inferences must it admit?
\vskip .15truein
We can often usefully separate this $epistemological$ problem
 from the heuristic problem
of programming the search for useful inferences.
\vfill\eject
\centerline{\bf Areas of common sense knowledge:}

\bigskip
\itemb effects of actions and other events
\itemb relations between appearance and reality
\itemb objects and their parts
\itemb {\it is-a} hierarchies
\itemb knowledge and belief
\itemb vision as a dynamic process
\itemb communication as a process
\itemb natural kinds vs. arbitrary boundaries
\vfill\eject
\centerline{\bf Open questions about common sense:}
\itemb effects of concurrent events
\itemb approximate theories, granularity, levels of detail
\itemb Which facts can be used directly in pattern-action rules?
\itemb Which facts can be used directly as fragments of logic programs?
\itemb When must the same fact be used in various ways?
\itemb expression of heuristic information as facts
\itemb maintaining modularity
\itemb ambiguity tolerance
\itemb elaboration tolerance
\itemb What is the correct formalism for describing modifications to theories?

\vfill\eject
\centerline{\bf The Advice Taker (1958):}

\noindent Represent as sentences of mathematical logic

\itemb general facts about the effects of actions and events

\itemb goals to be obtained

\itemb the principle that actions that achieve goals should be done

\itemb facts about the particular situation

Deduce a sentence of the form  $(SHOULD <action>)$ and then do the action.
\vfill\eject
The original Advice Taker plan won't work.  Why it won't work wasn't
apparent until something better came along - namely non-monotonic reasoning.

\vfill\eject
\centerline {\bf Formalized non-monotonic reasoning (1978):}

\noindent Monotonicity:

If a sentence {\it p} is inferred from a set {\it A} of premisses
and $A ⊂ B$, then {\it p} is inferred from $B$.

Commons sense reasoning includes non-monotonic steps.

\itemb Non-monotonic logic (McDermott and Doyle)

\itemb Default logic (Reiter)

\itemb Circumscription (McCarthy)
\vfill\eject
\centerline {\bf Formula Circumscription:}

\itemb $A(P)$ is an axiom about a predicate vector $P$.

\itemb $E(P,x)$ is a formula which we want to make false
for a maximal set of values of $x$.

$$A'(P) ≡ A(P) ∧ ∀P'.(A(P') ∧ (∀x.E(P',x) ⊃ E(P,x))$$
$$⊃ (∀x.E(P,x) ≡ E(P',x)))$$

\vfill\eject
\centerline {\bf Formula Circumscription:}

\itemb $A(P)$ is an axiom about a predicate vector $P$.

\itemb $E(P,x)$ is a formula which we want to make false
for a maximal set of values of $x$.

$$A'(P) ≡ A(P) ∧ ∀P'.(A(P') ∧ (∀x.E(P',x) ⊃ E(P,x))$$
$$⊃ (∀x.E(P,x) ≡ E(P',x)))$$

\vfill\eject
\centerline {\bf Formula Circumscription:}

\itemb $A(P)$ is an axiom about a predicate vector $P$.

\itemb $E(P,x)$ is a formula which we want to make false
for a maximal set of values of $x$.

$$A'(P) ≡ A(P) ∧ ∀P'.(A(P') ∧ (∀x.E(P',x) ⊃ E(P,x))$$
$$⊃ (∀x.E(P,x) ≡ E(P',x)))$$

\vfill\eject
\centerline {\bf Formula Circumscription:}

\itemb $A(P)$ is an axiom about a predicate vector $P$.

\itemb $E(P,x)$ is a formula which we want to make false
for a maximal set of values of $x$.

$$A'(P) ≡ A(P) ∧ ∀P'.(A(P') ∧ (∀x.E(P',x) ⊃ E(P,x))$$
$$⊃ (∀x.E(P,x) ≡ E(P',x)))$$

\vfill\eject
\centerline{\bf Common sense data base:}

	Once we have a good common sense reasoning
system and know how to express common sense knowledge,
we can create a common sense database usable by any
program that requires common sense.  Opinions differ
by a factor of 1000 as to how many facts the database
would need to be useful.  I think ten thousand, but some
think ten million.

\itemb Douglas Lenat's Cyc project is plunging ahead.

\itemb Edward Feigenbaum sometimes says that 10 million ordinary
expert system rules will do it.

\vfill\eject
\end