<s>
In	O
artificial	B-Application
intelligence	I-Application
,	O
symbolic	B-General_Concept
artificial	I-General_Concept
intelligence	I-General_Concept
is	O
the	O
term	O
for	O
the	O
collection	O
of	O
all	O
methods	O
in	O
artificial	B-Application
intelligence	I-Application
research	I-Application
that	O
are	O
based	O
on	O
high-level	O
symbolic	O
(	O
human-readable	O
)	O
representations	O
of	O
problems	O
,	O
logic	O
and	O
search	B-Application
.	O
</s>
<s>
Symbolic	B-General_Concept
AI	I-General_Concept
used	O
tools	O
such	O
as	O
logic	B-Language
programming	I-Language
,	O
production	B-Application
rules	I-Application
,	O
semantic	O
nets	O
and	O
frames	O
,	O
and	O
it	O
developed	O
applications	O
such	O
as	O
knowledge-based	B-General_Concept
systems	I-General_Concept
(	O
in	O
particular	O
,	O
expert	B-Application
systems	I-Application
)	O
,	O
symbolic	B-Algorithm
mathematics	I-Algorithm
,	O
automated	B-Application
theorem	I-Application
provers	I-Application
,	O
ontologies	B-Language
,	O
the	O
semantic	O
web	O
,	O
and	O
automated	B-Application
planning	I-Application
and	I-Application
scheduling	I-Application
systems	O
.	O
</s>
<s>
The	O
Symbolic	B-General_Concept
AI	I-General_Concept
paradigm	O
led	O
to	O
seminal	O
ideas	O
in	O
search	B-Application
,	O
symbolic	O
programming	O
languages	O
,	O
agents	B-General_Concept
,	O
multi-agent	O
systems	O
,	O
the	O
semantic	O
web	O
,	O
and	O
the	O
strengths	O
and	O
limitations	O
of	O
formal	O
knowledge	O
and	O
reasoning	O
systems	O
.	O
</s>
<s>
Symbolic	B-General_Concept
AI	I-General_Concept
was	O
the	O
dominant	O
paradigm	O
of	O
AI	B-Application
research	I-Application
from	O
the	O
mid-1950s	O
until	O
the	O
middle	O
1990s	O
.	O
</s>
<s>
An	O
early	O
boom	O
,	O
with	O
early	O
successes	O
such	O
as	O
the	O
Logic	B-Application
Theorist	I-Application
and	O
Samuel	O
's	O
Checker	O
's	O
Playing	O
Program	O
led	O
to	O
unrealistic	O
expectations	O
and	O
promises	O
and	O
was	O
followed	O
by	O
the	O
First	O
AI	B-Application
Winter	O
as	O
funding	O
dried	O
up	O
.	O
</s>
<s>
A	O
second	O
boom	O
(	O
1969	O
–	O
1986	O
)	O
occurred	O
with	O
the	O
rise	O
of	O
expert	B-Application
systems	I-Application
,	O
their	O
promise	O
of	O
capturing	O
corporate	O
expertise	O
,	O
and	O
an	O
enthusiastic	O
corporate	O
embrace	O
.	O
</s>
<s>
That	O
boom	O
,	O
and	O
some	O
early	O
successes	O
,	O
e.g.	O
,	O
with	O
XCON	B-Application
at	O
DEC	O
,	O
was	O
followed	O
again	O
by	O
later	O
disappointment	O
.	O
</s>
<s>
Another	O
,	O
second	O
,	O
AI	B-Application
Winter	O
(	O
1988	O
–	O
2011	O
)	O
followed	O
.	O
</s>
<s>
Subsequently	O
,	O
AI	B-Application
researchers	O
focused	O
on	O
addressing	O
underlying	O
problems	O
in	O
handling	O
uncertainty	O
and	O
in	O
knowledge	O
acquisition	O
.	O
</s>
<s>
Uncertainty	O
was	O
addressed	O
with	O
formal	O
methods	O
such	O
as	O
hidden	O
Markov	O
models	O
,	O
Bayesian	O
reasoning	O
,	O
and	O
statistical	B-Algorithm
relational	I-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Symbolic	O
machine	O
learning	O
addressed	O
the	O
knowledge	O
acquisition	O
problem	O
with	O
contributions	O
including	O
Version	B-General_Concept
Space	I-General_Concept
,	O
Valiant	O
's	O
PAC	O
learning	O
,	O
Quinlan	O
's	O
ID3	B-Algorithm
decision-tree	B-Algorithm
learning	O
,	O
case-based	O
learning	O
,	O
and	O
inductive	B-Application
logic	I-Application
programming	I-Application
to	O
learn	O
relations	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
,	O
a	O
subsymbolic	B-General_Concept
approach	O
,	O
had	O
been	O
pursued	O
from	O
early	O
days	O
and	O
was	O
to	O
reemerge	O
strongly	O
in	O
2012	O
.	O
</s>
<s>
Early	O
examples	O
are	O
Rosenblatt	O
's	O
perceptron	B-Algorithm
learning	O
work	O
,	O
the	O
backpropagation	B-Algorithm
work	O
of	O
Rumelhart	O
,	O
Hinton	O
and	O
Williams	O
,	O
and	O
work	O
in	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
by	O
LeCun	O
et	O
al	O
.	O
</s>
<s>
However	O
,	O
neural	B-Architecture
networks	I-Architecture
were	O
not	O
viewed	O
as	O
successful	O
until	O
about	O
2012	O
:	O
"	O
Until	O
Big	O
Data	O
became	O
commonplace	O
,	O
the	O
general	O
consensus	O
in	O
the	O
Al	O
community	O
was	O
that	O
the	O
so-called	O
neural-network	O
approach	O
was	O
hopeless	O
.	O
</s>
<s>
...	O
A	O
revolution	O
came	O
in	O
2012	O
,	O
when	O
a	O
number	O
of	O
people	O
,	O
including	O
a	O
team	O
of	O
researchers	O
working	O
with	O
Hinton	O
,	O
worked	O
out	O
a	O
way	O
to	O
use	O
the	O
power	O
of	O
GPUs	B-Architecture
to	O
enormously	O
increase	O
the	O
power	O
of	O
neural	O
networks.	O
"	O
</s>
<s>
Over	O
the	O
next	O
several	O
years	O
,	O
deep	B-Algorithm
learning	I-Algorithm
had	O
spectacular	O
success	O
in	O
handling	O
vision	B-Application
,	O
speech	O
recognition	O
,	O
speech	O
synthesis	O
,	O
image	O
generation	O
,	O
and	O
machine	O
translation	O
.	O
</s>
<s>
However	O
,	O
since	O
2020	O
,	O
as	O
inherent	O
difficulties	O
with	O
bias	O
,	O
explanation	O
,	O
comprehensibility	O
,	O
and	O
robustness	O
became	O
more	O
apparent	O
with	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
;	O
an	O
increasing	O
number	O
of	O
AI	B-Application
researchers	O
have	O
called	O
for	O
combining	O
the	O
best	O
of	O
both	O
the	O
symbolic	O
and	O
neural	B-Architecture
network	I-Architecture
approaches	O
and	O
addressing	O
areas	O
that	O
both	O
approaches	O
have	O
difficulty	O
with	O
,	O
such	O
as	O
common-sense	O
reasoning	O
.	O
</s>
<s>
Later	O
,	O
practitioners	O
using	O
knowledge-based	B-General_Concept
approaches	O
adopted	O
a	O
second	O
maxim	O
:	O
</s>
<s>
Finally	O
,	O
with	O
the	O
rise	O
of	O
deep	B-Algorithm
learning	I-Algorithm
,	O
the	O
symbolic	B-General_Concept
AI	I-General_Concept
approach	O
has	O
been	O
compared	O
to	O
deep	B-Algorithm
learning	I-Algorithm
as	O
complementary	O
"	O
with	O
...	O
parallels	O
having	O
been	O
drawn	O
many	O
times	O
by	O
AI	B-Application
researchers	O
between	O
Kahneman	O
's	O
research	O
on	O
human	O
reasoning	O
and	O
decision	O
making	O
–	O
reflected	O
in	O
his	O
book	O
Thinking	O
,	O
Fast	O
and	O
Slow	O
–	O
and	O
the	O
so-called	O
"	O
AI	B-Application
systems	O
1	O
and	O
2	O
"	O
,	O
which	O
would	O
in	O
principle	O
be	O
modelled	O
by	O
deep	B-Algorithm
learning	I-Algorithm
and	O
symbolic	B-Algorithm
reasoning	I-Algorithm
,	O
respectively.	O
"	O
</s>
<s>
In	O
this	O
view	O
,	O
symbolic	B-Algorithm
reasoning	I-Algorithm
is	O
more	O
apt	O
for	O
deliberative	O
reasoning	O
,	O
planning	B-Application
,	O
and	O
explanation	O
while	O
deep	B-Algorithm
learning	I-Algorithm
is	O
more	O
apt	O
for	O
fast	O
pattern	O
recognition	O
in	O
perceptual	O
applications	O
with	O
noisy	O
data	O
.	O
</s>
<s>
A	O
short	O
history	O
of	O
symbolic	B-General_Concept
AI	I-General_Concept
to	O
the	O
present	O
day	O
follows	O
below	O
.	O
</s>
<s>
Time	O
periods	O
and	O
titles	O
are	O
drawn	O
from	O
Henry	O
Kautz	O
's	O
2020	O
AAAI	O
Robert	O
S	O
.	O
Engelmore	O
Memorial	O
Lecture	O
and	O
the	O
longer	O
Wikipedia	O
article	O
on	O
the	O
History	O
of	O
AI	B-Application
,	O
with	O
dates	O
and	O
titles	O
differing	O
slightly	O
for	O
increased	O
clarity	O
.	O
</s>
<s>
Success	O
at	O
early	O
attempts	O
in	O
AI	B-Application
occurred	O
in	O
three	O
main	O
areas	O
:	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
knowledge	O
representation	O
,	O
and	O
heuristic	B-Algorithm
search	B-Application
,	O
contributing	O
to	O
high	O
expectations	O
.	O
</s>
<s>
This	O
section	O
summarizes	O
Kautz	O
's	O
reprise	O
of	O
early	O
AI	B-Application
history	O
.	O
</s>
<s>
A	O
robotic	O
turtle	O
,	O
with	O
sensors	O
,	O
motors	O
for	O
driving	O
and	O
steering	O
,	O
and	O
seven	O
vacuum	O
tubes	O
for	O
control	O
,	O
based	O
on	O
a	O
preprogrammed	O
neural	B-Architecture
net	I-Architecture
,	O
was	O
built	O
as	O
early	O
as	O
1948	O
.	O
</s>
<s>
This	O
work	O
can	O
be	O
seen	O
as	O
an	O
early	O
precursor	O
to	O
later	O
work	O
in	O
neural	B-Architecture
networks	I-Architecture
,	O
reinforcement	O
learning	O
,	O
and	O
situated	O
robotics	O
.	O
</s>
<s>
An	O
important	O
early	O
symbolic	B-General_Concept
AI	I-General_Concept
program	O
was	O
the	O
Logic	B-Application
theorist	I-Application
,	O
written	O
by	O
Allen	O
Newell	O
,	O
Herbert	O
Simon	O
and	O
Cliff	O
Shaw	O
in	O
1955	O
–	O
56	O
,	O
as	O
it	O
was	O
able	O
to	O
prove	O
38	O
elementary	O
theorems	O
from	O
Whitehead	O
and	O
Russell	O
's	O
Principia	O
Mathematica	O
.	O
</s>
<s>
GPS	O
solved	O
problems	O
represented	O
with	O
formal	O
operators	O
via	O
state-space	O
search	B-Application
using	O
means-ends	B-General_Concept
analysis	I-General_Concept
.	O
</s>
<s>
During	O
the	O
1960s	O
,	O
symbolic	O
approaches	O
achieved	O
great	O
success	O
at	O
simulating	O
intelligent	O
behavior	O
in	O
structured	O
environments	O
such	O
as	O
game-playing	O
,	O
symbolic	B-Algorithm
mathematics	I-Algorithm
,	O
and	O
theorem-proving	O
.	O
</s>
<s>
AI	B-Application
research	I-Application
was	O
centered	O
in	O
three	O
institutions	O
in	O
the	O
1960s	O
:	O
Carnegie	O
Mellon	O
University	O
,	O
Stanford	O
,	O
MIT	O
and	O
(	O
later	O
)	O
University	O
of	O
Edinburgh	O
.	O
</s>
<s>
Earlier	O
approaches	O
based	O
on	O
cybernetics	O
or	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
were	O
abandoned	O
or	O
pushed	O
into	O
the	O
background	O
.	O
</s>
<s>
Herbert	O
Simon	O
and	O
Allen	O
Newell	O
studied	O
human	O
problem-solving	O
skills	O
and	O
attempted	O
to	O
formalize	O
them	O
,	O
and	O
their	O
work	O
laid	O
the	O
foundations	O
of	O
the	O
field	O
of	O
artificial	B-Application
intelligence	I-Application
,	O
as	O
well	O
as	O
cognitive	O
science	O
,	O
operations	O
research	O
and	O
management	O
science	O
.	O
</s>
<s>
This	O
tradition	O
,	O
centered	O
at	O
Carnegie	O
Mellon	O
University	O
would	O
eventually	O
culminate	O
in	O
the	O
development	O
of	O
the	O
Soar	B-Architecture
architecture	O
in	O
the	O
middle	O
1980s	O
.	O
</s>
<s>
In	O
addition	O
to	O
the	O
highly-specialized	O
domain-specific	O
kinds	O
of	O
knowledge	O
that	O
we	O
will	O
see	O
later	O
used	O
in	O
expert	B-Application
systems	I-Application
,	O
early	O
symbolic	B-General_Concept
AI	I-General_Concept
researchers	O
discovered	O
another	O
more	O
general	O
application	O
of	O
knowledge	O
.	O
</s>
<s>
These	O
were	O
called	O
heuristics	B-Algorithm
,	O
rules	O
of	O
thumb	O
that	O
guide	O
a	O
search	B-Application
in	O
promising	O
directions	O
:	O
"	O
How	O
can	O
non-enumerative	O
search	B-Application
be	O
practical	O
when	O
the	O
underlying	O
problem	O
is	O
exponentially	O
hard	O
?	O
</s>
<s>
The	O
approach	O
advocated	O
by	O
Simon	O
and	O
Newell	O
is	O
to	O
employ	O
heuristics	B-Algorithm
:	O
fast	O
algorithms	O
that	O
may	O
fail	O
on	O
some	O
inputs	O
or	O
output	O
suboptimal	O
solutions.	O
"	O
</s>
<s>
Another	O
important	O
advance	O
was	O
to	O
find	O
a	O
way	O
to	O
apply	O
these	O
heuristics	B-Algorithm
that	O
guarantees	O
a	O
solution	O
will	O
be	O
found	O
,	O
if	O
there	O
is	O
one	O
,	O
not	O
withstanding	O
the	O
occasional	O
fallibility	O
of	O
heuristics	B-Algorithm
:	O
"	O
The	O
A*	B-Protocol
algorithm	I-Protocol
provided	O
a	O
general	O
frame	O
for	O
complete	O
and	O
optimal	O
heuristically	O
guided	O
search	B-Application
.	O
</s>
<s>
A*	B-Protocol
is	O
used	O
as	O
a	O
subroutine	O
within	O
practically	O
every	O
AI	B-Application
algorithm	O
today	O
but	O
is	O
still	O
no	O
magic	O
bullet	O
;	O
its	O
guarantee	O
of	O
completeness	O
is	O
bought	O
at	O
the	O
cost	O
of	O
worst-case	O
exponential	O
time	O
.	O
</s>
<s>
His	O
laboratory	O
at	O
Stanford	O
(	O
SAIL	O
)	O
focused	O
on	O
using	O
formal	O
logic	O
to	O
solve	O
a	O
wide	O
variety	O
of	O
problems	O
,	O
including	O
knowledge	O
representation	O
,	O
planning	B-Application
and	O
learning	O
.	O
</s>
<s>
Logic	O
was	O
also	O
the	O
focus	O
of	O
the	O
work	O
at	O
the	O
University	O
of	O
Edinburgh	O
and	O
elsewhere	O
in	O
Europe	O
which	O
led	O
to	O
the	O
development	O
of	O
the	O
programming	O
language	O
Prolog	B-Language
and	O
the	O
science	O
of	O
logic	B-Language
programming	I-Language
.	O
</s>
<s>
Researchers	O
at	O
MIT	O
(	O
such	O
as	O
Marvin	O
Minsky	O
and	O
Seymour	O
Papert	O
)	O
found	O
that	O
solving	O
difficult	O
problems	O
in	O
vision	B-Application
and	O
natural	B-Language
language	I-Language
processing	I-Language
required	O
ad	O
hoc	O
solutions	O
—	O
they	O
argued	O
that	O
no	O
simple	O
and	O
general	O
principle	O
(	O
like	O
logic	O
)	O
would	O
capture	O
all	O
the	O
aspects	O
of	O
intelligent	O
behavior	O
.	O
</s>
<s>
Commonsense	B-General_Concept
knowledge	I-General_Concept
bases	I-General_Concept
(	O
such	O
as	O
Doug	O
Lenat	O
's	O
Cyc	B-Language
)	O
are	O
an	O
example	O
of	O
"	O
scruffy	O
"	O
AI	B-Application
,	O
since	O
they	O
must	O
be	O
built	O
by	O
hand	O
,	O
one	O
complicated	O
concept	O
at	O
a	O
time	O
.	O
</s>
<s>
The	O
first	O
AI	B-Application
winter	O
was	O
a	O
shock	O
:	O
</s>
<s>
As	O
limitations	O
with	O
weak	O
,	O
domain-independent	O
methods	O
became	O
more	O
and	O
more	O
apparent	O
,	O
researchers	O
from	O
all	O
three	O
traditions	O
began	O
to	O
build	O
knowledge	O
into	O
AI	B-Application
applications	O
.	O
</s>
<s>
The	O
knowledge	O
revolution	O
was	O
driven	O
by	O
the	O
realization	O
that	O
knowledge	O
underlies	O
high-performance	O
,	O
domain-specific	O
AI	B-Application
applications	O
.	O
</s>
<s>
This	O
"	O
knowledge	O
revolution	O
"	O
led	O
to	O
the	O
development	O
and	O
deployment	O
of	O
expert	B-Application
systems	I-Application
(	O
introduced	O
by	O
Edward	O
Feigenbaum	O
)	O
,	O
the	O
first	O
commercially	O
successful	O
form	O
of	O
AI	B-Application
software	I-Application
.	O
</s>
<s>
Key	O
expert	B-Application
systems	I-Application
were	O
:	O
</s>
<s>
XCON	B-Application
,	O
to	O
configure	O
VAX	O
computers	O
,	O
a	O
then	O
laborious	O
process	O
that	O
could	O
take	O
up	O
to	O
90	O
days	O
.	O
</s>
<s>
XCON	B-Application
reduced	O
the	O
time	O
to	O
about	O
90	O
minutes	O
.	O
</s>
<s>
DENDRAL	O
is	O
considered	O
the	O
first	O
expert	B-Application
system	I-Application
that	O
relied	O
on	O
knowledge-intensive	O
problem-solving	O
.	O
</s>
<s>
The	O
other	O
expert	B-Application
systems	I-Application
mentioned	O
above	O
came	O
after	O
DENDRAL	O
.	O
</s>
<s>
MYCIN	O
exemplifies	O
the	O
classic	O
expert	B-Application
system	I-Application
architecture	O
of	O
a	O
knowledge-base	O
of	O
rules	O
coupled	O
to	O
a	O
symbolic	B-Algorithm
reasoning	I-Algorithm
mechanism	O
,	O
including	O
the	O
use	O
of	O
certainty	O
factors	O
to	O
handle	O
uncertainty	O
.	O
</s>
<s>
GUIDON	O
shows	O
how	O
an	O
explicit	O
knowledge	O
base	O
can	O
be	O
repurposed	O
for	O
a	O
second	O
application	O
,	O
tutoring	O
,	O
and	O
is	O
an	O
example	O
of	O
an	O
intelligent	O
tutoring	O
system	O
,	O
a	O
particular	O
kind	O
of	O
knowledge-based	B-General_Concept
application	O
.	O
</s>
<s>
XCON	B-Application
is	O
significant	O
because	O
of	O
the	O
millions	O
of	O
dollars	O
it	O
saved	O
DEC	O
,	O
which	O
triggered	O
the	O
expert	B-Application
system	I-Application
boom	O
where	O
most	O
all	O
major	O
corporations	O
in	O
the	O
US	O
had	O
expert	B-Application
systems	I-Application
groups	O
,	O
with	O
the	O
aim	O
to	O
capture	O
corporate	O
expertise	O
,	O
preserve	O
it	O
,	O
and	O
automate	O
it	O
:	O
</s>
<s>
Chess	B-Application
expert	O
knowledge	O
was	O
encoded	O
in	O
Deep	B-General_Concept
Blue	I-General_Concept
.	O
</s>
<s>
In	O
1996	O
,	O
this	O
allowed	O
IBM	O
's	O
Deep	B-General_Concept
Blue	I-General_Concept
,	O
with	O
the	O
help	O
of	O
symbolic	B-General_Concept
AI	I-General_Concept
,	O
to	O
win	O
in	O
a	O
game	O
of	O
chess	B-Application
against	O
the	O
world	O
champion	O
at	O
that	O
time	O
,	O
Garry	O
Kasparov	O
.	O
</s>
<s>
A	O
key	O
component	O
of	O
the	O
system	O
architecture	O
for	O
all	O
expert	B-Application
systems	I-Application
is	O
the	O
knowledge	O
base	O
,	O
which	O
stores	O
facts	O
and	O
rules	O
for	O
problem-solving	O
.	O
</s>
<s>
The	O
simplest	O
approach	O
for	O
an	O
expert	B-Application
system	I-Application
knowledge	O
base	O
is	O
simply	O
a	O
collection	O
or	O
network	O
of	O
production	B-Application
rules	I-Application
.	O
</s>
<s>
Production	B-Application
rules	I-Application
connect	O
symbols	O
in	O
a	O
relationship	O
similar	O
to	O
an	O
If-Then	B-Language
statement	I-Language
.	O
</s>
<s>
The	O
expert	B-Application
system	I-Application
processes	O
the	O
rules	O
to	O
make	O
deductions	O
and	O
to	O
determine	O
what	O
additional	O
information	O
it	O
needs	O
,	O
i.e.	O
</s>
<s>
For	O
example	O
,	O
OPS5	B-Language
,	O
CLIPS	B-Application
and	O
their	O
successors	O
Jess	B-Language
and	O
Drools	B-Language
operate	O
in	O
this	O
fashion	O
.	O
</s>
<s>
Expert	B-Application
systems	I-Application
can	O
operate	O
in	O
either	O
a	O
forward	O
chaining	O
–	O
from	O
evidence	O
to	O
conclusions	O
–	O
or	O
backward	O
chaining	O
–	O
from	O
goals	O
to	O
needed	O
data	O
and	O
prerequisites	O
–	O
manner	O
.	O
</s>
<s>
More	O
advanced	O
knowledge-based	B-General_Concept
systems	I-General_Concept
,	O
such	O
as	O
Soar	B-Architecture
can	O
also	O
perform	O
meta-level	O
reasoning	O
,	O
that	O
is	O
reasoning	O
about	O
their	O
own	O
reasoning	O
in	O
terms	O
of	O
deciding	O
how	O
to	O
solve	O
problems	O
and	O
monitoring	O
the	O
success	O
of	O
problem-solving	O
strategies	O
.	O
</s>
<s>
Blackboard	B-General_Concept
systems	I-General_Concept
are	O
a	O
second	O
kind	O
of	O
knowledge-based	B-General_Concept
or	O
expert	B-Application
system	I-Application
architecture	O
.	O
</s>
<s>
One	O
example	O
,	O
the	O
BB1	O
blackboard	B-General_Concept
architecture	O
was	O
originally	O
inspired	O
by	O
studies	O
of	O
how	O
humans	O
plan	O
to	O
perform	O
multiple	O
tasks	O
in	O
a	O
trip	O
.	O
</s>
<s>
An	O
innovation	O
of	O
BB1	O
was	O
to	O
apply	O
the	O
same	O
blackboard	B-General_Concept
model	I-General_Concept
to	O
solving	O
its	O
own	O
control	O
problem	O
,	O
i.e.	O
,	O
its	O
controller	O
performed	O
meta-level	O
reasoning	O
with	O
knowledge	O
sources	O
that	O
monitored	O
how	O
well	O
a	O
plan	O
or	O
the	O
problem-solving	O
was	O
proceeding	O
,	O
and	O
could	O
switch	O
from	O
one	O
strategy	O
to	O
another	O
as	O
conditions	O
–	O
such	O
as	O
goals	O
or	O
times	O
–	O
changed	O
.	O
</s>
<s>
BB1	O
was	O
applied	O
in	O
multiple	O
domains	O
:	O
construction	O
site	O
planning	B-Application
,	O
intelligent	O
tutoring	O
systems	O
,	O
and	O
real-time	O
patient	O
monitoring	O
.	O
</s>
<s>
At	O
the	O
height	O
of	O
the	O
AI	B-Application
boom	O
,	O
companies	O
such	O
as	O
Symbolics	O
,	O
LMI	B-Operating_System
,	O
and	O
Texas	O
Instruments	O
were	O
selling	O
LISP	B-Operating_System
machines	I-Operating_System
specifically	O
targeted	O
to	O
accelerate	O
the	O
development	O
of	O
AI	B-Application
applications	O
and	O
research	O
.	O
</s>
<s>
In	O
addition	O
,	O
several	O
artificial	B-Application
intelligence	I-Application
companies	O
,	O
such	O
as	O
Teknowledge	O
and	O
Inference	O
Corporation	O
,	O
were	O
selling	O
expert	B-Application
system	I-Application
shells	O
,	O
training	O
,	O
and	O
consulting	O
to	O
corporations	O
.	O
</s>
<s>
Unfortunately	O
,	O
the	O
AI	B-Application
boom	O
did	O
not	O
last	O
and	O
Kautz	O
best	O
describes	O
the	O
second	O
AI	B-Application
winter	O
that	O
followed	O
:	O
</s>
<s>
and	O
Bayesian	O
approaches	O
were	O
applied	O
successfully	O
in	O
expert	B-Application
systems	I-Application
.	O
</s>
<s>
Even	O
later	O
,	O
in	O
the	O
1990s	O
,	O
statistical	B-Algorithm
relational	I-Algorithm
learning	I-Algorithm
,	O
an	O
approach	O
that	O
combines	O
probability	O
with	O
logical	O
formulas	O
,	O
allowed	O
probability	O
to	O
be	O
combined	O
with	O
first-order	O
logic	O
,	O
e.g.	O
,	O
with	O
either	O
Markov	O
Logic	O
Networks	O
or	O
Probabilistic	O
Soft	O
Logic	O
.	O
</s>
<s>
In	O
contrast	O
to	O
the	O
knowledge-intensive	O
approach	O
of	O
Meta-DENDRAL	O
,	O
Ross	O
Quinlan	O
invented	O
a	O
domain-independent	O
approach	O
to	O
statistical	O
classification	O
,	O
decision	B-Algorithm
tree	I-Algorithm
learning	I-Algorithm
,	O
starting	O
first	O
with	O
ID3	B-Algorithm
and	O
then	O
later	O
extending	O
its	O
capabilities	O
to	O
C4.5	B-Algorithm
.	O
</s>
<s>
The	O
decision	B-Algorithm
trees	I-Algorithm
created	O
are	O
glass	B-General_Concept
box	I-General_Concept
,	O
interpretable	O
classifiers	O
,	O
with	O
human-interpretable	O
classification	O
rules	O
.	O
</s>
<s>
Tom	O
Mitchell	O
introduced	O
version	B-General_Concept
space	I-General_Concept
learning	I-General_Concept
which	O
describes	O
learning	O
as	O
search	B-Application
through	O
a	O
space	O
of	O
hypotheses	O
,	O
with	O
upper	O
,	O
more	O
general	O
,	O
and	O
lower	O
,	O
more	O
specific	O
,	O
boundaries	O
encompassing	O
all	O
viable	O
hypotheses	O
consistent	O
with	O
the	O
examples	O
seen	O
so	O
far	O
.	O
</s>
<s>
E.g.	O
,	O
John	O
Anderson	O
provided	O
a	O
cognitive	O
model	O
of	O
human	O
learning	O
where	O
skill	O
practice	O
results	O
in	O
a	O
compilation	O
of	O
rules	O
from	O
a	O
declarative	B-Language
format	O
to	O
a	O
procedural	O
format	O
with	O
his	O
ACT-R	B-Language
cognitive	B-Architecture
architecture	I-Architecture
.	O
</s>
<s>
ACT-R	B-Language
has	O
been	O
used	O
successfully	O
to	O
model	O
aspects	O
of	O
human	O
cognition	O
,	O
such	O
as	O
learning	O
and	O
retention	O
.	O
</s>
<s>
ACT-R	B-Language
is	O
also	O
used	O
in	O
intelligent	O
tutoring	O
systems	O
,	O
called	O
cognitive	B-Application
tutors	I-Application
,	O
to	O
successfully	O
teach	O
geometry	O
,	O
computer	O
programming	O
,	O
and	O
algebra	O
to	O
school	O
children	O
.	O
</s>
<s>
Inductive	B-Application
logic	I-Application
programming	I-Application
was	O
another	O
approach	O
to	O
learning	O
that	O
allowed	O
logic	B-Language
programs	I-Language
to	O
be	O
synthesized	O
from	O
input-output	O
examples	O
.	O
</s>
<s>
E.g.	O
,	O
Ehud	O
Shapiro	O
's	O
MIS	O
(	O
Model	O
Inference	O
System	O
)	O
could	O
synthesize	O
Prolog	B-Language
programs	O
from	O
examples	O
.	O
</s>
<s>
John	O
R	O
.	O
Koza	O
applied	O
genetic	B-Algorithm
algorithms	I-Algorithm
to	O
program	B-Application
synthesis	I-Application
to	O
create	O
genetic	B-Algorithm
programming	I-Algorithm
,	O
which	O
he	O
used	O
to	O
synthesize	O
LISP	B-Language
programs	O
.	O
</s>
<s>
Finally	O
,	O
Manna	O
and	O
Waldinger	O
provided	O
a	O
more	O
general	O
approach	O
to	O
program	B-Application
synthesis	I-Application
that	O
synthesizes	O
a	O
functional	B-Language
program	I-Language
in	O
the	O
course	O
of	O
proving	O
its	O
specifications	O
to	O
be	O
correct	O
.	O
</s>
<s>
Another	O
alternative	O
to	O
logic	O
,	O
genetic	B-Algorithm
algorithms	I-Algorithm
and	O
genetic	B-Algorithm
programming	I-Algorithm
are	O
based	O
on	O
an	O
evolutionary	O
model	O
of	O
learning	O
,	O
where	O
sets	O
of	O
rules	O
are	O
encoded	O
into	O
populations	O
,	O
the	O
rules	O
govern	O
the	O
behavior	O
of	O
individuals	O
,	O
and	O
selection	O
of	O
the	O
fittest	O
prunes	O
out	O
sets	O
of	O
unsuitable	O
rules	O
over	O
many	O
generations	O
.	O
</s>
<s>
Symbolic	O
machine	O
learning	O
was	O
applied	O
to	O
learning	O
concepts	O
,	O
rules	O
,	O
heuristics	B-Algorithm
,	O
and	O
problem-solving	O
.	O
</s>
<s>
Doug	O
Lenat	O
's	O
Eurisko	B-Algorithm
,	O
for	O
example	O
,	O
learned	O
heuristics	B-Algorithm
to	O
beat	O
human	O
players	O
at	O
the	O
Traveller	O
role-playing	O
game	O
for	O
two	O
years	O
in	O
a	O
row	O
.	O
</s>
<s>
Neuro-symbolic	B-General_Concept
AI	I-General_Concept
attempts	O
to	O
integrate	O
neural	O
and	O
symbolic	O
architectures	O
in	O
a	O
manner	O
that	O
addresses	O
strengths	O
and	O
weaknesses	O
of	O
each	O
,	O
in	O
a	O
complementary	O
fashion	O
,	O
in	O
order	O
to	O
support	O
robust	O
AI	B-Application
capable	O
of	O
reasoning	O
,	O
learning	O
,	O
and	O
cognitive	O
modeling	O
.	O
</s>
<s>
As	O
argued	O
by	O
Valiant	O
and	O
many	O
others	O
,	O
the	O
effective	O
construction	O
of	O
rich	O
computational	O
cognitive	O
models	O
demands	O
the	O
combination	O
of	O
sound	O
symbolic	B-Algorithm
reasoning	I-Algorithm
and	O
efficient	O
(	O
machine	O
)	O
learning	O
models	O
.	O
</s>
<s>
"	O
To	O
build	O
a	O
robust	O
,	O
knowledge-driven	O
approach	O
to	O
AI	B-Application
we	O
must	O
have	O
the	O
machinery	O
of	O
symbol-manipulation	O
in	O
our	O
toolkit	O
.	O
</s>
<s>
System	O
1	O
is	O
the	O
kind	O
used	O
for	O
pattern	O
recognition	O
while	O
System	O
2	O
is	O
far	O
better	O
suited	O
for	O
planning	B-Application
,	O
deduction	O
,	O
and	O
deliberative	O
thinking	O
.	O
</s>
<s>
In	O
this	O
view	O
,	O
deep	B-Algorithm
learning	I-Algorithm
best	O
models	O
the	O
first	O
kind	O
of	O
thinking	O
while	O
symbolic	B-Algorithm
reasoning	I-Algorithm
best	O
models	O
the	O
second	O
kind	O
and	O
both	O
are	O
needed	O
.	O
</s>
<s>
A	O
series	O
of	O
workshops	O
on	O
neuro-symbolic	B-General_Concept
reasoning	O
has	O
been	O
held	O
every	O
year	O
since	O
2005	O
,	O
see	O
for	O
details	O
.	O
</s>
<s>
Henry	O
Kautz	O
's	O
taxonomy	O
of	O
neuro-symbolic	B-General_Concept
architectures	O
,	O
along	O
with	O
some	O
examples	O
,	O
follows	O
:	O
</s>
<s>
Symbolic	O
Neural	O
symbolic	O
—	O
is	O
the	O
current	O
approach	O
of	O
many	O
neural	O
models	O
in	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
where	O
words	O
or	O
subword	O
tokens	O
are	O
both	O
the	O
ultimate	O
input	O
and	O
output	O
of	O
large	O
language	O
models	O
.	O
</s>
<s>
Examples	O
include	O
BERT	B-General_Concept
,	O
RoBERTa	O
,	O
and	O
GPT-3	B-General_Concept
.	O
</s>
<s>
Symbolic[Neural]—is	O
exemplified	O
by	O
AlphaGo	B-Application
,	O
where	O
symbolic	O
techniques	O
are	O
used	O
to	O
call	O
neural	O
techniques	O
.	O
</s>
<s>
In	O
this	O
case	O
the	O
symbolic	O
approach	O
is	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
and	O
the	O
neural	O
techniques	O
learn	O
how	O
to	O
evaluate	O
game	O
positions	O
.	O
</s>
<s>
Neural:Symbolic	O
→	O
Neural	O
—	O
relies	O
on	O
symbolic	B-Algorithm
reasoning	I-Algorithm
to	O
generate	O
or	O
label	O
training	O
data	O
that	O
is	O
subsequently	O
learned	O
by	O
a	O
deep	B-Algorithm
learning	I-Algorithm
model	O
,	O
e.g.	O
,	O
to	O
train	O
a	O
neural	O
model	O
for	O
symbolic	B-Algorithm
computation	I-Algorithm
by	O
using	O
a	O
Macsyma-like	O
symbolic	B-Algorithm
mathematics	I-Algorithm
system	O
to	O
create	O
or	O
label	O
examples	O
.	O
</s>
<s>
Neural_{Symbolic}—uses	O
a	O
neural	B-Architecture
net	I-Architecture
that	O
is	O
generated	O
from	O
symbolic	O
rules	O
.	O
</s>
<s>
An	O
example	O
is	O
the	O
Neural	O
Theorem	O
Prover	O
,	O
which	O
constructs	O
a	O
neural	B-Architecture
network	I-Architecture
from	O
an	O
AND-OR	B-Data_Structure
proof	O
tree	O
generated	O
from	O
knowledge	O
base	O
rules	O
and	O
terms	O
.	O
</s>
<s>
Neural[Symbolic]—allows	O
a	O
neural	O
model	O
to	O
directly	O
call	O
a	O
symbolic	B-Algorithm
reasoning	I-Algorithm
engine	O
,	O
e.g.	O
,	O
to	O
perform	O
an	O
action	O
or	O
evaluate	O
a	O
state	O
.	O
</s>
<s>
How	O
should	O
symbolic	O
structures	O
be	O
represented	O
within	O
neural	B-Architecture
networks	I-Architecture
and	O
extracted	O
from	O
them	O
?	O
</s>
<s>
The	O
key	O
AI	B-Application
programming	O
language	O
in	O
the	O
US	O
during	O
the	O
last	O
symbolic	B-General_Concept
AI	I-General_Concept
boom	O
period	O
was	O
LISP	B-Language
.	O
</s>
<s>
LISP	B-Language
is	O
the	O
second	O
oldest	O
programming	O
language	O
after	O
FORTRAN	B-Application
and	O
was	O
created	O
in	O
1958	O
by	O
John	O
McCarthy	O
.	O
</s>
<s>
LISP	B-Language
provided	O
the	O
first	O
read-eval-print	B-Application
loop	I-Application
to	O
support	O
rapid	O
program	O
development	O
.	O
</s>
<s>
It	O
had	O
the	O
first	O
self-hosting	O
compiler	O
,	O
meaning	O
that	O
the	O
compiler	O
itself	O
was	O
originally	O
written	O
in	O
LISP	B-Language
and	O
then	O
ran	O
interpretively	O
to	O
compile	O
the	O
compiler	O
code	O
.	O
</s>
<s>
Other	O
key	O
innovations	O
pioneered	O
by	O
LISP	B-Language
that	O
have	O
spread	O
to	O
other	O
programming	O
languages	O
include	O
:	O
</s>
<s>
In	O
contrast	O
to	O
the	O
US	O
,	O
in	O
Europe	O
the	O
key	O
AI	B-Application
programming	O
language	O
during	O
that	O
same	O
period	O
was	O
Prolog	B-Language
.	O
</s>
<s>
Prolog	B-Language
provided	O
a	O
built-in	O
store	O
of	O
facts	O
and	O
clauses	O
that	O
could	O
be	O
queried	O
by	O
a	O
read-eval-print	B-Application
loop	I-Application
.	O
</s>
<s>
As	O
a	O
subset	O
of	O
first-order	O
logic	O
Prolog	B-Language
was	O
based	O
on	O
Horn	B-Application
clauses	I-Application
with	O
a	O
closed-world	B-Application
assumption	I-Application
--	O
any	O
facts	O
not	O
known	O
were	O
considered	O
false	O
--	O
and	O
a	O
unique	O
name	O
assumption	O
for	O
primitive	O
terms	O
--	O
e.g.	O
,	O
the	O
identifier	O
barack_obama	O
was	O
considered	O
to	O
refer	O
to	O
exactly	O
one	O
object	O
.	O
</s>
<s>
Backtracking	B-Algorithm
and	O
unification	B-Algorithm
are	O
built-in	O
to	O
Prolog	B-Language
.	O
</s>
<s>
Alain	O
Colmerauer	O
and	O
Philippe	O
Roussel	O
are	O
credited	O
as	O
the	O
inventors	O
of	O
Prolog	B-Language
.	O
</s>
<s>
Prolog	B-Language
is	O
a	O
form	O
of	O
logic	B-Language
programming	I-Language
,	O
which	O
was	O
invented	O
by	O
Robert	O
Kowalski	O
.	O
</s>
<s>
Its	O
history	O
was	O
also	O
influenced	O
by	O
Carl	O
Hewitt	O
's	O
PLANNER	B-Application
,	O
an	O
assertional	O
database	O
with	O
pattern-directed	O
invocation	O
of	O
methods	O
.	O
</s>
<s>
For	O
more	O
detail	O
see	O
the	O
section	O
on	O
the	O
origins	O
of	O
Prolog	B-Language
in	O
the	O
PLANNER	B-Application
article	O
.	O
</s>
<s>
Prolog	B-Language
is	O
also	O
a	O
kind	O
of	O
declarative	B-Language
programming	I-Language
.	O
</s>
<s>
No	O
explicit	O
series	O
of	O
actions	O
is	O
required	O
,	O
as	O
is	O
the	O
case	O
with	O
imperative	B-Application
programming	I-Application
languages	I-Application
.	O
</s>
<s>
Japan	O
championed	O
Prolog	B-Language
for	O
its	O
Fifth	B-Device
Generation	I-Device
Project	I-Device
,	O
intending	O
to	O
build	O
special	O
hardware	O
for	O
high	O
performance	O
.	O
</s>
<s>
Similarly	O
,	O
LISP	B-Operating_System
machines	I-Operating_System
were	O
built	O
to	O
run	O
LISP	B-Language
,	O
but	O
as	O
the	O
second	O
AI	B-Application
boom	O
turned	O
to	O
bust	O
these	O
companies	O
could	O
not	O
compete	O
with	O
new	O
workstations	O
that	O
could	O
now	O
run	O
LISP	B-Language
or	O
Prolog	B-Language
natively	O
at	O
comparable	O
speeds	O
.	O
</s>
<s>
Smalltalk	B-Application
was	O
another	O
influential	O
AI	B-Application
programming	O
language	O
.	O
</s>
<s>
For	O
example	O
it	O
introduced	O
metaclasses	B-Language
and	O
,	O
along	O
with	O
Flavors	B-Language
and	O
CommonLoops	B-Language
,	O
influenced	O
the	O
Common	B-Application
Lisp	I-Application
Object	I-Application
System	I-Application
,	O
or	O
(	O
CLOS	B-Application
)	O
,	O
that	O
is	O
now	O
part	O
of	O
Common	B-Language
Lisp	I-Language
,	O
the	O
current	O
standard	O
Lisp	B-Language
dialect	O
.	O
</s>
<s>
CLOS	B-Application
is	O
a	O
Lisp-based	O
object-oriented	B-Language
system	O
that	O
allows	O
multiple	B-Application
inheritance	I-Application
,	O
in	O
addition	O
to	O
incremental	O
extensions	O
to	O
both	O
classes	O
and	O
metaclasses	B-Language
,	O
thus	O
providing	O
a	O
run-time	O
meta-object	O
protocol	O
.	O
</s>
<s>
For	O
other	O
AI	B-Application
programming	O
languages	O
see	O
this	O
list	B-Language
of	I-Language
programming	I-Language
languages	I-Language
for	I-Language
artificial	I-Language
intelligence	I-Language
.	O
</s>
<s>
Currently	O
,	O
Python	B-Language
,	O
a	O
multi-paradigm	O
programming	O
language	O
,	O
is	O
the	O
most	O
popular	O
programming	O
language	O
,	O
partly	O
due	O
to	O
its	O
extensive	O
package	O
library	O
that	O
supports	O
data	O
science	O
,	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Python	B-Language
includes	O
a	O
read-eval-print	B-Application
loop	I-Application
,	O
functional	O
elements	O
such	O
as	O
higher-order	B-Language
functions	I-Language
,	O
and	O
object-oriented	B-Language
programming	I-Language
that	O
includes	O
metaclasses	B-Language
.	O
</s>
<s>
Search	B-Application
arises	O
in	O
many	O
kinds	O
of	O
problem	O
solving	O
,	O
including	O
planning	B-Application
,	O
constraint	B-Application
satisfaction	I-Application
,	O
and	O
playing	O
games	O
such	O
as	O
checkers	O
,	O
chess	B-Application
,	O
and	O
go	O
.	O
</s>
<s>
The	O
best	O
known	O
AI-search	O
tree	O
search	B-Application
algorithms	I-Application
are	O
breadth-first	B-Algorithm
search	I-Algorithm
,	O
depth-first	B-Algorithm
search	I-Algorithm
,	O
A*	B-Protocol
,	O
and	O
Monte	B-Application
Carlo	I-Application
Search	I-Application
.	O
</s>
<s>
Key	O
search	B-Application
algorithms	I-Application
for	O
Boolean	B-Algorithm
satisfiability	I-Algorithm
are	O
WalkSAT	B-Application
,	O
conflict-driven	B-Application
clause	I-Application
learning	I-Application
,	O
and	O
the	O
DPLL	B-Application
algorithm	I-Application
.	O
</s>
<s>
For	O
adversarial	O
search	B-Application
when	O
playing	O
games	O
,	O
alpha-beta	B-Algorithm
pruning	I-Algorithm
,	O
branch	B-Algorithm
and	I-Algorithm
bound	I-Algorithm
,	O
and	O
minimax	B-Algorithm
were	O
early	O
contributions	O
.	O
</s>
<s>
Ontologies	B-Language
model	O
key	O
concepts	O
and	O
their	O
relationships	O
in	O
a	O
domain	O
.	O
</s>
<s>
Example	O
ontologies	B-Language
are	O
YAGO	O
,	O
WordNet	B-General_Concept
,	O
and	O
DOLCE	O
.	O
</s>
<s>
DOLCE	O
is	O
an	O
example	O
of	O
an	O
upper	O
ontology	B-Language
that	O
can	O
be	O
used	O
for	O
any	O
domain	O
while	O
WordNet	B-General_Concept
is	O
a	O
lexical	O
resource	O
that	O
can	O
also	O
be	O
viewed	O
as	O
an	O
ontology	B-Language
.	O
</s>
<s>
YAGO	O
incorporates	O
WordNet	B-General_Concept
as	O
part	O
of	O
its	O
ontology	B-Language
,	O
to	O
align	O
facts	O
extracted	O
from	O
Wikipedia	O
with	O
WordNet	B-General_Concept
synsets	B-General_Concept
.	O
</s>
<s>
The	O
Disease	O
Ontology	B-Language
is	O
an	O
example	O
of	O
a	O
medical	O
ontology	B-Language
currently	O
being	O
used	O
.	O
</s>
<s>
Description	B-Language
logic	I-Language
is	O
a	O
logic	O
for	O
automated	O
classification	O
of	O
ontologies	B-Language
and	O
for	O
detecting	O
inconsistent	O
classification	O
data	O
.	O
</s>
<s>
OWL	B-Language
is	O
a	O
language	O
used	O
to	O
represent	O
ontologies	B-Language
with	O
description	B-Language
logic	I-Language
.	O
</s>
<s>
Protégé	B-Language
is	O
a	O
ontology	B-Language
editor	O
that	O
can	O
read	O
in	O
OWL	B-Language
ontologies	I-Language
and	O
then	O
check	O
consistency	O
with	O
deductive	B-Algorithm
classifiers	I-Algorithm
such	O
as	O
such	O
as	O
HermiT	O
.	O
</s>
<s>
First-order	O
logic	O
is	O
more	O
general	O
than	O
description	B-Language
logic	I-Language
.	O
</s>
<s>
The	O
automated	B-Application
theorem	I-Application
provers	I-Application
discussed	O
below	O
can	O
prove	O
theorems	O
in	O
first-order	O
logic	O
.	O
</s>
<s>
Horn	B-Application
clause	I-Application
logic	O
is	O
more	O
restricted	O
than	O
first-order	O
logic	O
and	O
is	O
used	O
in	O
logic	B-Language
programming	I-Language
languages	I-Language
such	O
as	O
Prolog	B-Language
.	O
</s>
<s>
Examples	O
of	O
automated	B-Application
theorem	I-Application
provers	I-Application
for	O
first-order	O
logic	O
are	O
:	O
</s>
<s>
Prover9	B-Application
can	O
be	O
used	O
in	O
conjunction	O
with	O
the	O
Mace4	B-Application
model	B-Application
checker	I-Application
.	O
</s>
<s>
ACL2	B-Application
is	O
a	O
theorem	O
prover	O
that	O
can	O
handle	O
proofs	O
by	O
induction	O
and	O
is	O
a	O
descendant	O
of	O
the	O
Boyer-Moore	B-Application
Theorem	I-Application
Prover	I-Application
,	O
also	O
known	O
as	O
Nqthm	B-Application
.	O
</s>
<s>
Knowledge-based	B-General_Concept
systems	I-General_Concept
have	O
an	O
explicit	O
knowledge	O
base	O
,	O
typically	O
of	O
rules	O
,	O
to	O
enhance	O
reusability	O
across	O
domains	O
by	O
separating	O
procedural	O
code	O
and	O
domain	O
knowledge	O
.	O
</s>
<s>
Forward	O
chaining	O
inference	O
engines	O
are	O
the	O
most	O
common	O
,	O
and	O
are	O
seen	O
in	O
CLIPS	B-Application
and	O
OPS5	B-Language
.	O
</s>
<s>
Backward	O
chaining	O
occurs	O
in	O
Prolog	B-Language
,	O
where	O
a	O
more	O
limited	O
logical	O
representation	O
is	O
used	O
,	O
Horn	B-Application
Clauses	I-Application
.	O
</s>
<s>
Pattern-matching	O
,	O
specifically	O
unification	B-Algorithm
,	O
is	O
used	O
in	O
Prolog	B-Language
.	O
</s>
<s>
This	O
kind	O
of	O
meta-level	O
reasoning	O
is	O
used	O
in	O
Soar	B-Architecture
and	O
in	O
the	O
BB1	O
blackboard	B-General_Concept
architecture	O
.	O
</s>
<s>
Cognitive	B-Architecture
architectures	I-Architecture
such	O
as	O
ACT-R	B-Language
may	O
have	O
additional	O
capabilities	O
,	O
such	O
as	O
the	O
ability	O
to	O
compile	O
frequently	O
used	O
knowledge	O
into	O
higher-level	O
chunks	O
.	O
</s>
<s>
Cyc	B-Language
has	O
attempted	O
to	O
capture	O
useful	O
common-sense	O
knowledge	O
and	O
has	O
"	O
micro-theories	O
"	O
to	O
handle	O
particular	O
kinds	O
of	O
domain-specific	O
reasoning	O
.	O
</s>
<s>
Similarly	O
,	O
Allen	O
's	O
temporal	B-Application
interval	I-Application
algebra	I-Application
is	O
a	O
simplification	O
of	O
reasoning	O
about	O
time	O
and	O
Region	B-Application
Connection	I-Application
Calculus	I-Application
is	O
a	O
simplification	O
of	O
reasoning	O
about	O
spatial	O
relationships	O
.	O
</s>
<s>
Both	O
can	O
be	O
solved	O
with	O
constraint	B-Application
solvers	I-Application
.	O
</s>
<s>
Constraint	B-Application
solvers	I-Application
perform	O
a	O
more	O
limited	O
kind	O
of	O
inference	O
than	O
first-order	O
logic	O
.	O
</s>
<s>
They	O
can	O
simplify	O
sets	O
of	O
spatiotemporal	O
constraints	O
,	O
such	O
as	O
those	O
for	O
RCC	B-Application
or	O
Temporal	B-Application
Algebra	I-Application
,	O
along	O
with	O
solving	O
other	O
kinds	O
of	O
puzzle	O
problems	O
,	O
such	O
as	O
Wordle	B-Application
,	O
Sudoku	O
,	O
cryptarithmetic	O
problems	O
,	O
and	O
so	O
on	O
.	O
</s>
<s>
Constraint	B-Application
logic	I-Application
programming	I-Application
can	O
be	O
used	O
to	O
solve	O
scheduling	O
problems	O
,	O
for	O
example	O
with	O
constraint	B-Application
handling	I-Application
rules	I-Application
(	O
CHR	O
)	O
.	O
</s>
<s>
The	O
General	O
Problem	O
Solver	O
(	O
GPS	O
)	O
cast	O
planning	B-Application
as	O
problem-solving	O
used	O
means-ends	B-General_Concept
analysis	I-General_Concept
to	O
create	O
plans	O
.	O
</s>
<s>
STRIPS	B-Application
took	O
a	O
different	O
approach	O
,	O
viewing	O
planning	B-Application
as	O
theorem	B-Application
proving	I-Application
.	O
</s>
<s>
Graphplan	B-Application
takes	O
a	O
least-commitment	O
approach	O
to	O
planning	B-Application
,	O
rather	O
than	O
sequentially	O
choosing	O
actions	O
from	O
an	O
initial	O
state	O
,	O
working	O
forwards	O
,	O
or	O
a	O
goal	O
state	O
if	O
working	O
backwards	O
.	O
</s>
<s>
Satplan	B-Application
is	O
an	O
approach	O
to	O
planning	B-Application
where	O
a	O
planning	B-Application
problem	O
is	O
reduced	O
to	O
a	O
Boolean	B-Algorithm
satisfiability	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
Natural	B-Language
language	I-Language
processing	I-Language
focuses	O
on	O
treating	O
language	O
as	O
data	O
to	O
perform	O
tasks	O
such	O
as	O
identifying	O
topics	O
without	O
necessarily	O
understanding	O
the	O
intended	O
meaning	O
.	O
</s>
<s>
Parsing	B-Language
,	O
tokenizing	O
,	O
spelling	B-Application
correction	I-Application
,	O
part-of-speech	O
tagging	O
,	O
noun	B-General_Concept
and	I-General_Concept
verb	I-General_Concept
phrase	I-General_Concept
chunking	I-General_Concept
are	O
all	O
aspects	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
long	O
handled	O
by	O
symbolic	B-General_Concept
AI	I-General_Concept
,	O
but	O
since	O
improved	O
by	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
.	O
</s>
<s>
In	O
symbolic	B-General_Concept
AI	I-General_Concept
,	O
discourse	O
representation	O
theory	O
and	O
first-order	O
logic	O
have	O
been	O
used	O
to	O
represent	O
sentence	O
meanings	O
.	O
</s>
<s>
Latent	O
semantic	O
analysis	O
(	O
LSA	O
)	O
and	O
explicit	B-General_Concept
semantic	I-General_Concept
analysis	I-General_Concept
also	O
provided	O
vector	O
representations	O
of	O
documents	O
.	O
</s>
<s>
New	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
based	O
on	O
Transformer	B-Algorithm
models	I-Algorithm
have	O
now	O
eclipsed	O
these	O
earlier	O
symbolic	B-General_Concept
AI	I-General_Concept
approaches	O
and	O
attained	O
state-of-the-art	O
performance	O
in	O
natural	B-Language
language	I-Language
processing	I-Language
.	O
</s>
<s>
However	O
,	O
Transformer	B-Algorithm
models	I-Algorithm
are	O
opaque	O
and	O
do	O
not	O
yet	O
produce	O
human-interpretable	O
semantic	O
representations	O
for	O
sentences	O
and	O
documents	O
.	O
</s>
<s>
Agents	B-General_Concept
are	O
autonomous	O
systems	O
embedded	O
in	O
an	O
environment	O
they	O
perceive	O
and	O
act	O
upon	O
in	O
some	O
sense	O
.	O
</s>
<s>
Russell	O
and	O
Norvig	O
's	O
standard	O
textbook	O
on	O
artificial	B-Application
intelligence	I-Application
is	O
organized	O
to	O
reflect	O
agent	O
architectures	O
of	O
increasing	O
sophistication	O
.	O
</s>
<s>
The	O
sophistication	O
of	O
agents	B-General_Concept
varies	O
from	O
simple	O
reactive	O
agents	B-General_Concept
,	O
to	O
those	O
with	O
a	O
model	O
of	O
the	O
world	O
and	O
automated	B-Application
planning	I-Application
capabilities	O
,	O
possibly	O
a	O
BDI	O
agent	O
,	O
i.e.	O
,	O
one	O
with	O
beliefs	O
,	O
desires	O
,	O
and	O
intentions	O
–	O
or	O
alternatively	O
a	O
reinforcement	O
learning	O
model	O
learned	O
over	O
time	O
to	O
choose	O
actions	O
–	O
up	O
to	O
a	O
combination	O
of	O
alternative	O
architectures	O
,	O
such	O
as	O
a	O
neuro-symbolic	B-General_Concept
architecture	O
that	O
includes	O
deep	B-Algorithm
learning	I-Algorithm
for	O
perception	O
.	O
</s>
<s>
In	O
contrast	O
,	O
a	O
multi-agent	O
system	O
consists	O
of	O
multiple	O
agents	B-General_Concept
that	O
communicate	O
amongst	O
themselves	O
with	O
some	O
inter-agent	O
communication	O
language	O
such	O
as	O
Knowledge	B-Language
Query	I-Language
and	I-Language
Manipulation	I-Language
Language	I-Language
(	O
KQML	B-Language
)	O
.	O
</s>
<s>
The	O
agents	B-General_Concept
need	O
not	O
all	O
have	O
the	O
same	O
internal	O
architecture	O
.	O
</s>
<s>
Advantages	O
of	O
multi-agent	O
systems	O
include	O
the	O
ability	O
to	O
divide	O
work	O
among	O
the	O
agents	B-General_Concept
and	O
to	O
increase	O
fault	O
tolerance	O
when	O
agents	B-General_Concept
are	O
lost	O
.	O
</s>
<s>
Research	O
problems	O
include	O
how	B-Application
agents	I-Application
reach	I-Application
consensus	I-Application
,	O
distributed	B-Operating_System
problem	I-Operating_System
solving	I-Operating_System
,	O
multi-agent	B-Algorithm
learning	I-Algorithm
,	O
multi-agent	O
planning	B-Application
,	O
and	O
distributed	B-Application
constraint	I-Application
optimization	I-Application
.	O
</s>
<s>
Controversies	O
arose	O
from	O
early	O
on	O
in	O
symbolic	B-General_Concept
AI	I-General_Concept
,	O
both	O
within	O
the	O
field	O
—	O
e.g.	O
,	O
between	O
logicists	O
(	O
the	O
pro-logic	O
"	O
neats	O
"	O
)	O
and	O
non-logicists	O
(	O
the	O
anti-logic	O
"	O
scruffies	O
"	O
)	O
—	O
and	O
between	O
those	O
who	O
embraced	O
AI	B-Application
but	O
rejected	O
symbolic	O
approaches	O
—	O
primarily	O
connectionists	O
—	O
and	O
those	O
outside	O
the	O
field	O
.	O
</s>
<s>
Critiques	O
from	O
outside	O
of	O
the	O
field	O
were	O
primarily	O
from	O
philosophers	O
,	O
on	O
intellectual	O
grounds	O
,	O
but	O
also	O
from	O
funding	O
agencies	O
,	O
especially	O
during	O
the	O
two	O
AI	B-Application
winters	O
.	O
</s>
<s>
Connectionist	O
approaches	O
include	O
earlier	O
work	O
on	O
neural	B-Architecture
networks	I-Architecture
,	O
such	O
as	O
perceptrons	B-Algorithm
;	O
work	O
in	O
the	O
mid	O
to	O
late	O
80s	O
,	O
such	O
as	O
Danny	O
Hillis	O
's	O
Connection	B-Device
Machine	I-Device
and	O
Yann	O
Le	O
Cun	O
's	O
advances	O
in	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
;	O
to	O
today	O
's	O
more	O
advanced	O
approaches	O
,	O
such	O
as	O
Transformers	B-Algorithm
,	O
GANs	B-Algorithm
,	O
and	O
other	O
work	O
in	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Implementationism	O
—	O
where	O
connectionist	O
architectures	O
implement	O
the	O
capabilities	O
for	O
symbolic	B-Algorithm
processing	I-Algorithm
,	O
</s>
<s>
Radical	O
connectionism	O
—	O
where	O
symbolic	B-Algorithm
processing	I-Algorithm
is	O
rejected	O
totally	O
,	O
and	O
connectionist	O
architectures	O
underlie	O
intelligence	O
and	O
are	O
fully	O
sufficient	O
to	O
explain	O
it	O
,	O
</s>
<s>
Moderate	O
connectionism	O
—	O
where	O
symbolic	B-Algorithm
processing	I-Algorithm
and	O
connectionist	O
architectures	O
are	O
viewed	O
as	O
complementary	O
and	O
both	O
are	O
required	O
for	O
intelligence	O
.	O
</s>
<s>
Olazaran	O
,	O
in	O
his	O
sociological	O
history	O
of	O
the	O
controversies	O
within	O
the	O
neural	B-Architecture
network	I-Architecture
community	O
,	O
described	O
the	O
moderate	O
connectionism	O
view	O
as	O
essentially	O
compatible	O
with	O
current	O
research	O
in	O
neuro-symbolic	O
hybrids:The	O
third	O
and	O
last	O
position	O
I	O
would	O
like	O
to	O
examine	O
here	O
is	O
what	O
I	O
call	O
the	O
moderate	O
connectionist	O
view	O
,	O
a	O
more	O
eclectic	O
view	O
of	O
the	O
current	O
debate	O
between	O
connectionism	O
and	O
symbolic	B-General_Concept
AI	I-General_Concept
.	O
</s>
<s>
Gary	O
Marcus	O
has	O
claimed	O
that	O
the	O
animus	O
in	O
the	O
deep	B-Algorithm
learning	I-Algorithm
community	O
against	O
symbolic	O
approaches	O
now	O
may	O
be	O
more	O
sociological	O
than	O
philosophical:To	O
think	O
that	O
we	O
can	O
simply	O
abandon	O
symbol-manipulation	O
is	O
to	O
suspend	O
disbelief	O
.	O
</s>
<s>
And	O
yet	O
,	O
for	O
the	O
most	O
part	O
,	O
that	O
's	O
how	O
most	O
current	O
AI	B-Application
proceeds	O
.	O
</s>
<s>
The	O
deep	B-Algorithm
learning	I-Algorithm
hope	O
—	O
seemingly	O
grounded	O
not	O
so	O
much	O
in	O
science	O
,	O
but	O
in	O
a	O
sort	O
of	O
historical	O
grudge	O
—	O
is	O
that	O
intelligent	O
behavior	O
will	O
emerge	O
purely	O
from	O
the	O
confluence	O
of	O
massive	O
data	O
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Where	O
classical	O
computers	O
and	O
software	O
solve	O
tasks	O
by	O
defining	O
sets	O
of	O
symbol-manipulating	O
rules	O
dedicated	O
to	O
particular	O
jobs	O
,	O
such	O
as	O
editing	O
a	O
line	O
in	O
a	O
word	O
processor	O
or	O
performing	O
a	O
calculation	O
in	O
a	O
spreadsheet	O
,	O
neural	B-Architecture
networks	I-Architecture
typically	O
try	O
to	O
solve	O
tasks	O
by	O
statistical	O
approximation	O
and	O
learning	O
from	O
examples.According	O
to	O
Marcus	O
,	O
Geoffrey	O
Hinton	O
and	O
his	O
colleagues	O
have	O
been	O
vehemently	O
"	O
anti-symbolic	O
"	O
:When	O
deep	B-Algorithm
learning	I-Algorithm
reemerged	O
in	O
2012	O
,	O
it	O
was	O
with	O
a	O
kind	O
of	O
take-no-prisoners	O
attitude	O
that	O
has	O
characterized	O
most	O
of	O
the	O
last	O
decade	O
.	O
</s>
<s>
He	O
gave	O
a	O
talk	O
at	O
an	O
AI	B-Application
workshop	O
at	O
Stanford	O
comparing	O
symbols	O
to	O
aether	O
,	O
one	O
of	O
science	O
's	O
greatest	O
mistakes	O
.	O
</s>
<s>
In	O
2016	O
,	O
Yann	O
LeCun	O
,	O
Bengio	O
,	O
and	O
Hinton	O
wrote	O
a	O
manifesto	O
for	O
deep	B-Algorithm
learning	I-Algorithm
in	O
one	O
of	O
science	O
's	O
most	O
important	O
journals	O
,	O
Nature	O
.	O
</s>
<s>
Turing	O
award	O
winner	O
Judea	O
Pearl	O
offers	O
a	O
critique	O
of	O
machine	O
learning	O
which	O
,	O
unfortunately	O
,	O
conflates	O
the	O
terms	O
machine	O
learning	O
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Similarly	O
,	O
when	O
Geoffrey	O
Hinton	O
refers	O
to	O
symbolic	B-General_Concept
AI	I-General_Concept
,	O
the	O
connotation	O
of	O
the	O
term	O
tends	O
to	O
be	O
that	O
of	O
expert	B-Application
systems	I-Application
dispossessed	O
of	O
any	O
ability	O
to	O
learn	O
.	O
</s>
<s>
Machine	O
learning	O
is	O
not	O
confined	O
to	O
association	B-Algorithm
rule	I-Algorithm
mining	I-Algorithm
,	O
c.f.	O
</s>
<s>
the	O
body	O
of	O
work	O
on	O
symbolic	O
ML	O
and	O
relational	B-Application
learning	I-Application
(	O
the	O
differences	O
to	O
deep	B-Algorithm
learning	I-Algorithm
being	O
the	O
choice	O
of	O
representation	O
,	O
localist	O
logical	O
rather	O
than	O
distributed	O
,	O
and	O
the	O
non-use	O
of	O
gradient-based	B-Algorithm
learning	I-Algorithm
algorithms	I-Algorithm
)	O
.	O
</s>
<s>
Equally	O
,	O
symbolic	B-General_Concept
AI	I-General_Concept
is	O
not	O
just	O
about	O
production	B-Application
rules	I-Application
written	O
by	O
hand	O
.	O
</s>
<s>
A	O
proper	O
definition	O
of	O
AI	B-Application
concerns	O
knowledge	O
representation	O
and	O
reasoning	O
,	O
autonomous	O
multi-agent	O
systems	O
,	O
planning	B-Application
and	O
argumentation	B-General_Concept
,	O
as	O
well	O
as	O
learning	O
.	O
</s>
<s>
Dreyfus	O
predicted	O
AI	B-Application
would	O
only	O
be	O
suitable	O
for	O
toy	O
problems	O
,	O
and	O
thought	O
that	O
building	O
more	O
complex	O
systems	O
or	O
scaling	O
up	O
the	O
idea	O
towards	O
useful	O
software	O
would	O
not	O
be	O
possible	O
.	O
</s>
<s>
John	O
Haugeland	O
,	O
another	O
philosopher	O
,	O
similarly	O
argued	O
against	O
rule-based	B-Language
symbolic	B-General_Concept
AI	I-General_Concept
in	O
his	O
book	O
Artificial	B-Application
Intelligence	I-Application
:	O
The	O
Very	O
Idea	O
,	O
calling	O
it	O
GOFAI	O
(	O
"	O
Good	B-General_Concept
Old-Fashioned	I-General_Concept
Artificial	I-General_Concept
Intelligence	I-General_Concept
"	O
)	O
.	O
</s>
<s>
Russell	O
and	O
Norvig	O
explain	O
that	O
these	O
arguments	O
were	O
targeted	O
to	O
the	O
symbolic	B-General_Concept
AI	I-General_Concept
of	O
the	O
1980s:The	O
technology	O
they	O
criticized	O
came	O
to	O
be	O
called	O
Good	B-General_Concept
Old-Fashioned	I-General_Concept
AI	I-General_Concept
(	O
GOFAI	O
)	O
.	O
</s>
<s>
Since	O
then	O
,	O
probabilistic	O
reasoning	O
systems	O
have	O
extended	O
the	O
capability	O
of	O
symbolic	B-General_Concept
AI	I-General_Concept
so	O
they	O
can	O
be	O
much	O
"	O
more	O
appropriate	O
for	O
open-ended	O
domains.	O
"	O
</s>
<s>
However	O
,	O
Dreyfus	O
raised	O
another	O
argument	O
that	O
cannot	O
be	O
addressed	O
by	O
disembodied	O
symbolic	B-General_Concept
AI	I-General_Concept
systems:One	O
of	O
Dreyfus	O
's	O
strongest	O
arguments	O
is	O
for	O
situated	O
agents	B-General_Concept
rather	O
than	O
disembodied	O
logical	O
inference	O
engines	O
.	O
</s>
<s>
Under	O
the	O
embodied	O
cognition	O
approach	O
,	O
robotics	O
,	O
vision	B-Application
,	O
and	O
other	O
sensors	O
become	O
central	O
,	O
not	O
peripheral	O
.	O
</s>
<s>
Rodney	O
Brooks	O
,	O
created	O
behavior-based	O
robotics	O
,	O
also	O
named	O
Nouvelle	B-General_Concept
AI	I-General_Concept
,	O
as	O
an	O
alternative	O
to	O
both	O
symbolic	B-General_Concept
AI	I-General_Concept
and	O
connectionist	O
AI	B-Application
.	O
</s>
<s>
Instead	O
,	O
he	O
created	O
the	O
subsumption	B-Architecture
architecture	I-Architecture
,	O
a	O
layered	O
architecture	O
for	O
embodied	O
agents	B-General_Concept
.	O
</s>
<s>
He	O
criticized	O
AI	B-Application
researchers	O
for	O
defining	O
AI	B-Application
problems	O
for	O
their	O
systems	O
,	O
when	O
:	O
"	O
There	O
is	O
no	O
clean	O
division	O
between	O
perception	O
(	O
abstraction	O
)	O
and	O
reasoning	O
in	O
the	O
real	O
world.	O
"	O
</s>
<s>
In	O
the	O
Nouvelle	B-General_Concept
AI	I-General_Concept
approach	O
,	O
"	O
First	O
,	O
it	O
is	O
vitally	O
important	O
to	O
test	O
the	O
Creatures	O
we	O
build	O
in	O
the	O
real	O
world	O
;	O
i.e.	O
,	O
in	O
the	O
same	O
world	O
that	O
we	O
humans	O
inhabit	O
.	O
</s>
<s>
His	O
emphasis	O
on	O
real-world	O
testing	O
was	O
in	O
contrast	O
to	O
"	O
Early	O
work	O
in	O
AI	B-Application
concentrated	O
on	O
games	O
,	O
geometrical	O
problems	O
,	O
symbolic	O
algebra	O
,	O
theorem	B-Application
proving	I-Application
,	O
and	O
other	O
formal	O
systems	O
"	O
and	O
the	O
use	O
of	O
the	O
blocks	B-General_Concept
world	I-General_Concept
in	O
symbolic	B-General_Concept
AI	I-General_Concept
systems	O
such	O
as	O
SHRDLU	O
.	O
</s>
<s>
Symbolic	B-General_Concept
AI	I-General_Concept
has	O
been	O
criticized	O
as	O
disembodied	O
,	O
liable	O
to	O
the	O
qualification	O
problem	O
,	O
and	O
poor	O
in	O
handling	O
the	O
perceptual	O
problems	O
where	O
deep	B-Algorithm
learning	I-Algorithm
excels	O
.	O
</s>
<s>
In	O
turn	O
,	O
connectionist	O
AI	B-Application
has	O
been	O
criticized	O
as	O
poorly	O
suited	O
for	O
deliberative	O
step-by-step	O
problem	O
solving	O
,	O
incorporating	O
knowledge	O
,	O
and	O
handling	O
planning	B-Application
.	O
</s>
<s>
Finally	O
,	O
Nouvelle	B-General_Concept
AI	I-General_Concept
excels	O
in	O
reactive	O
and	O
real-world	O
robotics	O
domains	O
but	O
has	O
been	O
criticized	O
for	O
difficulties	O
in	O
incorporating	O
learning	O
and	O
knowledge	O
.	O
</s>
<s>
Hybrid	B-General_Concept
AIs	I-General_Concept
incorporating	O
one	O
or	O
more	O
of	O
these	O
approaches	O
are	O
currently	O
viewed	O
as	O
the	O
path	O
forward	O
.	O
</s>
<s>
Russell	O
and	O
Norvig	O
conclude	O
that:Overall	O
,	O
Dreyfus	O
saw	O
areas	O
where	O
AI	B-Application
did	O
not	O
have	O
complete	O
answers	O
and	O
said	O
that	O
Al	O
is	O
therefore	O
impossible	O
;	O
we	O
now	O
see	O
many	O
of	O
these	O
same	O
areas	O
undergoing	O
continued	O
research	O
and	O
development	O
leading	O
to	O
increased	O
capability	O
,	O
not	O
impossibility	O
.	O
</s>
