<s>
Soar	B-Architecture
is	O
a	O
cognitive	B-Architecture
architecture	I-Architecture
,	O
originally	O
created	O
by	O
John	O
Laird	O
,	O
Allen	O
Newell	O
,	O
and	O
Paul	O
Rosenbloom	O
at	O
Carnegie	O
Mellon	O
University	O
.	O
</s>
<s>
The	O
goal	O
of	O
the	O
Soar	B-Architecture
project	O
is	O
to	O
develop	O
the	O
fixed	O
computational	O
building	O
blocks	O
necessary	O
for	O
general	O
intelligent	B-General_Concept
agents	I-General_Concept
–	O
agents	O
that	O
can	O
perform	O
a	O
wide	O
range	O
of	O
tasks	O
and	O
encode	O
,	O
use	O
,	O
and	O
learn	O
all	O
types	O
of	O
knowledge	O
to	O
realize	O
the	O
full	O
range	O
of	O
cognitive	O
capabilities	O
found	O
in	O
humans	O
,	O
such	O
as	O
decision	O
making	O
,	O
problem	O
solving	O
,	O
planning	O
,	O
and	O
natural-language	B-General_Concept
understanding	I-General_Concept
.	O
</s>
<s>
Since	O
its	O
beginnings	O
in	O
1983	O
as	O
John	O
Laird	O
’s	O
thesis	O
,	O
it	O
has	O
been	O
widely	O
used	O
by	O
AI	O
researchers	O
to	O
create	O
intelligent	B-General_Concept
agents	I-General_Concept
and	O
cognitive	O
models	O
of	O
different	O
aspects	O
of	O
human	O
behavior	O
.	O
</s>
<s>
The	O
most	O
current	O
and	O
comprehensive	O
description	O
of	O
Soar	B-Architecture
is	O
the	O
2012	O
book	O
,	O
The	O
Soar	B-Architecture
Cognitive	B-Architecture
Architecture	I-Architecture
.	O
</s>
<s>
Soar	B-Architecture
embodies	O
multiple	O
hypotheses	O
about	O
the	O
computational	O
structures	O
underlying	O
general	O
intelligence	O
,	O
many	O
of	O
which	O
are	O
shared	O
with	O
other	O
cognitive	B-Architecture
architectures	I-Architecture
,	O
including	O
ACT-R	B-Language
,	O
which	O
was	O
created	O
by	O
John	O
R	O
.	O
Anderson	O
,	O
and	O
LIDA	B-Architecture
,	O
which	O
was	O
created	O
by	O
Stan	O
Franklin	O
.	O
</s>
<s>
Recently	O
,	O
the	O
emphasis	O
on	O
Soar	B-Architecture
has	O
been	O
on	O
general	O
AI	O
(	O
functionality	O
and	O
efficiency	O
)	O
,	O
whereas	O
the	O
emphasis	O
on	O
ACT-R	B-Language
has	O
always	O
been	O
on	O
cognitive	O
modeling	O
(	O
detailed	O
modeling	O
of	O
human	O
cognition	O
)	O
.	O
</s>
<s>
The	O
original	O
theory	O
of	O
cognition	O
underlying	O
Soar	B-Architecture
is	O
the	O
Problem	O
Space	O
Hypothesis	O
,	O
which	O
is	O
described	O
in	O
Allen	O
Newell	O
's	O
book	O
,	O
Unified	O
Theories	O
of	O
Cognition	O
.	O
</s>
<s>
and	O
dates	O
back	O
to	O
one	O
of	O
the	O
first	O
AI	O
systems	O
created	O
,	O
Newell	O
,	O
Simon	O
,	O
and	O
Shaw	O
's	O
Logic	B-Application
Theorist	I-Application
,	O
first	O
presented	O
in	O
1955	O
.	O
</s>
<s>
(	O
Soar	B-Architecture
’s	O
name	O
is	O
derived	O
from	O
this	O
basic	O
cycle	O
of	O
State	O
,	O
Operator	O
,	O
And	O
Result	O
;	O
however	O
,	O
it	O
is	O
no	O
longer	O
regarded	O
as	O
an	O
acronym	O
.	O
)	O
</s>
<s>
A	O
second	O
hypothesis	O
of	O
Soar	B-Architecture
’s	O
theory	O
is	O
that	O
although	O
only	O
a	O
single	O
operator	O
can	O
be	O
selected	O
at	O
each	O
step	O
,	O
forcing	O
a	O
serial	O
bottleneck	O
,	O
the	O
processes	O
of	O
selection	O
and	O
application	O
are	O
implemented	O
through	O
parallel	O
rule	O
firings	O
,	O
which	O
provide	O
context-dependent	O
retrieval	O
of	O
procedural	O
knowledge	O
.	O
</s>
<s>
This	O
can	O
lead	O
to	O
a	O
stack	O
of	O
substates	O
,	O
where	O
traditional	O
problem	O
methods	O
,	O
such	O
as	O
planning	O
or	O
hierarchical	B-Application
task	I-Application
decomposition	I-Application
,	O
naturally	O
arise	O
.	O
</s>
<s>
A	O
fourth	O
hypothesis	O
in	O
Soar	B-Architecture
is	O
that	O
the	O
underlying	O
structure	O
is	O
modular	O
,	O
but	O
not	O
in	O
terms	O
of	O
task	O
or	O
capability	O
based	O
modules	O
,	O
such	O
as	O
planning	O
or	O
language	O
,	O
but	O
instead	O
as	O
task	O
independent	O
modules	O
including	O
:	O
a	O
decision	O
making	O
module	O
;	O
memory	O
modules	O
(	O
short-term	O
spatial/visual	O
and	O
working	O
memories	O
;	O
long-term	O
procedural	O
,	O
declarative	O
,	O
and	O
episodic	O
memories	O
)	O
,	O
learning	O
mechanisms	O
associated	O
with	O
all	O
long-term	O
memories	O
;	O
and	O
perceptual	O
and	O
motor	O
modules	O
.	O
</s>
<s>
The	O
hypothesis	O
that	O
a	O
symbolic	B-General_Concept
system	I-General_Concept
is	O
necessary	O
for	O
general	O
intelligence	O
is	O
known	O
as	O
the	O
physical	O
symbol	O
system	O
hypothesis	O
.	O
</s>
<s>
An	O
important	O
evolution	O
in	O
Soar	B-Architecture
is	O
that	O
all	O
symbolic	O
structures	O
have	O
associated	O
statistical	O
metadata	O
(	O
such	O
as	O
information	O
on	O
recency	O
and	O
frequency	O
of	O
use	O
,	O
or	O
expected	O
future	O
reward	O
)	O
that	O
influences	O
retrieval	O
,	O
maintenance	O
,	O
and	O
learning	O
of	O
the	O
symbolic	O
structures	O
.	O
</s>
<s>
Soar	B-Architecture
’s	O
main	O
processing	O
cycle	O
arises	O
from	O
the	O
interaction	O
between	O
procedural	O
memory	O
(	O
its	O
knowledge	O
about	O
how	O
to	O
do	O
things	O
)	O
and	O
working	O
memory	O
(	O
its	O
representation	O
of	O
the	O
current	O
situation	O
)	O
to	O
support	O
the	O
selection	O
and	O
application	O
of	O
operators	O
.	O
</s>
<s>
This	O
combination	O
of	O
rules	O
and	O
working	O
memory	O
is	O
also	O
called	O
a	O
production	B-Application
system	I-Application
.	O
</s>
<s>
In	O
contrast	O
to	O
most	O
production	B-Application
systems	I-Application
,	O
in	O
Soar	B-Architecture
,	O
all	O
rules	O
that	O
match	O
,	O
fire	O
in	O
parallel	O
.	O
</s>
<s>
Instead	O
of	O
having	O
the	O
selection	O
of	O
a	O
single	O
rule	O
being	O
the	O
crux	O
of	O
decision	O
making	O
,	O
Soar	B-Architecture
’s	O
decision	O
making	O
occurs	O
through	O
the	O
selection	O
and	O
applications	O
of	O
operators	O
,	O
that	O
are	O
proposed	O
,	O
evaluated	O
,	O
and	O
applied	O
by	O
rules	O
.	O
</s>
<s>
The	O
changes	O
to	O
working	O
memory	O
can	O
be	O
simple	O
inferences	O
,	O
queries	O
for	O
retrieval	O
from	O
Soar	B-Architecture
’s	O
long-term	O
semantic	O
or	O
episodic	O
memories	O
,	O
commands	O
to	O
the	O
motor	O
system	O
to	O
perform	O
actions	O
in	O
an	O
environment	O
,	O
or	O
interactions	O
with	O
the	O
Spatial	O
Visual	O
System	O
(	O
SVS	O
)	O
,	O
which	O
is	O
working	O
memory	O
’s	O
interface	O
to	O
perception	O
.	O
</s>
<s>
Soar	B-Architecture
supports	O
reinforcement	O
learning	O
,	O
which	O
tunes	O
the	O
values	O
of	O
rules	O
that	O
create	O
numeric	O
preferences	O
for	O
evaluating	O
operators	O
,	O
based	O
on	O
reward	O
.	O
</s>
<s>
Substates	O
provide	O
a	O
means	O
for	O
on-demand	O
complex	O
reasoning	O
,	O
including	O
hierarchical	B-Application
task	I-Application
decomposition	I-Application
,	O
planning	O
,	O
and	O
access	O
to	O
the	O
declarative	O
long-term	O
memories	O
.	O
</s>
<s>
Soar	B-Architecture
’s	O
chunking	O
mechanism	O
compiles	O
the	O
processing	O
in	O
the	O
substate	O
which	O
led	O
to	O
results	O
into	O
rules	O
.	O
</s>
<s>
Recently	O
,	O
the	O
overall	O
Universal	O
Subgoaling	O
procedure	O
has	O
been	O
extended	O
through	O
a	O
mechanism	O
of	O
goal-directed	O
and	O
automatic	O
knowledge	O
base	O
augmentation	O
that	O
allows	O
to	O
solve	O
an	O
impasse	O
by	O
recombining	O
,	O
in	O
an	O
innovative	O
and	O
problem-oriented	O
way	O
,	O
the	O
knowledge	O
possessed	O
by	O
a	O
Soar	B-Architecture
agent	O
.	O
</s>
<s>
To	O
support	O
interaction	O
with	O
vision	O
systems	O
and	O
non-symbolic	O
reasoning	O
,	O
Soar	B-Architecture
has	O
its	O
Spatial	O
Visual	O
System	O
(	O
SVS	O
)	O
.	O
</s>
<s>
A	O
Soar	B-Architecture
agent	O
using	O
SVS	O
can	O
create	O
filters	O
to	O
automatically	O
extract	O
features	O
and	O
relations	O
from	O
its	O
scene	O
graph	O
,	O
which	O
are	O
then	O
added	O
to	O
working	O
memory	O
.	O
</s>
<s>
In	O
addition	O
,	O
a	O
Soar	B-Architecture
agent	O
can	O
add	O
structures	O
to	O
SVS	O
and	O
use	O
it	O
for	O
mental	O
imagery	O
.	O
</s>
<s>
Semantic	O
Memory	O
(	O
SMEM	O
)	O
in	O
Soar	B-Architecture
is	O
designed	O
to	O
be	O
a	O
very	O
large	O
long-term	O
memory	O
of	O
fact-like	O
structures	O
.	O
</s>
<s>
SMEM	O
structures	O
have	O
activation	O
values	O
that	O
represent	O
the	O
frequency	O
or	O
recency	O
of	O
usage	O
of	O
each	O
memory	O
,	O
implementing	O
the	O
base-level	O
activation	O
scheme	O
originally	O
developed	O
for	O
ACT-R	B-Language
.	O
During	O
retrieval	O
,	O
the	O
structure	O
in	O
SMEM	O
that	O
matches	O
the	O
query	O
and	O
has	O
the	O
highest	O
activation	O
is	O
retrieved	O
.	O
</s>
<s>
Soar	B-Architecture
also	O
supports	O
spreading	B-Algorithm
activation	I-Algorithm
,	O
where	O
activation	O
spreads	O
from	O
SMEM	O
structures	O
that	O
have	O
been	O
retrieved	O
into	O
working	O
memory	O
to	O
other	O
long-term	O
memories	O
that	O
they	O
are	O
linked	O
to	O
.	O
</s>
<s>
Spreading	B-Algorithm
activation	I-Algorithm
is	O
a	O
mechanism	O
for	O
allowing	O
the	O
current	O
context	O
to	O
influence	O
retrievals	O
from	O
semantic	O
memory	O
.	O
</s>
<s>
Each	O
of	O
Soar	B-Architecture
’s	O
long-term	O
memories	O
have	O
associated	O
online	O
learning	O
mechanisms	O
that	O
create	O
new	O
structures	O
or	O
modify	O
metadata	O
based	O
on	O
an	O
agent	O
’s	O
experience	O
.	O
</s>
<s>
For	O
example	O
,	O
Soar	B-Architecture
learns	O
new	O
rules	O
for	O
procedural	O
memory	O
through	O
a	O
process	O
called	O
chunking	O
and	O
uses	O
reinforcement	O
learning	O
to	O
tune	O
rules	O
involved	O
in	O
the	O
selection	O
of	O
operators	O
.	O
</s>
<s>
The	O
standard	O
approach	O
to	O
developing	O
an	O
agent	O
in	O
Soar	B-Architecture
starts	O
with	O
writing	O
rules	O
that	O
are	O
loaded	O
into	O
procedural	O
memory	O
,	O
and	O
initializing	O
semantic	O
memory	O
with	O
appropriate	O
declarative	O
knowledge	O
.	O
</s>
<s>
The	O
process	O
of	O
agent	O
development	O
is	O
explained	O
in	O
detail	O
in	O
the	O
official	O
Soar	B-Architecture
manual	O
as	O
well	O
as	O
in	O
several	O
tutorials	O
which	O
are	O
provided	O
at	O
the	O
research	O
group	O
's	O
.	O
</s>
<s>
The	O
Soar	B-Architecture
architecture	O
is	O
maintained	O
and	O
extended	O
by	O
John	O
Laird	O
's	O
research	O
group	O
at	O
the	O
University	O
of	O
Michigan	O
.	O
</s>
<s>
Soar	B-Architecture
can	O
interface	O
with	O
external	O
language	O
environments	O
including	O
C++	O
,	O
Java	O
,	O
Tcl	O
,	O
and	O
Python	O
through	O
the	O
Soar	B-Architecture
Markup	O
Language	O
(	O
SML	O
)	O
.	O
</s>
<s>
SML	O
is	O
a	O
primary	O
mechanism	O
for	O
creating	O
instances	O
of	O
Soar	B-Architecture
agents	O
and	O
interacting	O
with	O
their	O
I/O	O
links	O
.	O
</s>
<s>
JSoar	O
is	O
an	O
implementation	O
of	O
Soar	B-Architecture
written	O
in	O
Java	O
.	O
</s>
<s>
Below	O
is	O
a	O
historical	O
list	O
of	O
different	O
areas	O
of	O
applications	O
that	O
have	O
been	O
implemented	O
in	O
Soar	B-Architecture
.	O
</s>
<s>
There	O
have	O
been	O
over	O
a	O
hundred	O
systems	O
implemented	O
in	O
Soar	B-Architecture
,	O
although	O
the	O
vast	O
majority	O
of	O
them	O
are	O
toy	O
tasks	O
or	O
puzzles	O
.	O
</s>
<s>
Throughout	O
its	O
history	O
,	O
Soar	B-Architecture
has	O
been	O
used	O
to	O
implement	O
a	O
wide	O
variety	O
of	O
classic	O
AI	O
puzzles	O
and	O
games	O
,	O
such	O
as	O
Tower	O
of	O
Hanoi	O
,	O
Water	O
Jug	O
,	O
Tic	O
Tac	O
Toe	O
,	O
Eight	O
Puzzle	O
,	O
Missionaries	O
and	O
Cannibals	O
,	O
and	O
variations	O
of	O
the	O
Blocks	B-General_Concept
world	I-General_Concept
.	O
</s>
<s>
One	O
of	O
the	O
initial	O
achievements	O
of	O
Soar	B-Architecture
was	O
showing	O
that	O
many	O
different	O
weak	O
methods	O
would	O
naturally	O
arise	O
from	O
the	O
task	O
knowledge	O
that	O
was	O
encoded	O
in	O
it	O
,	O
a	O
property	O
called	O
,	O
the	O
Universal	O
Weak	O
Method	O
.	O
</s>
<s>
The	O
first	O
large-scale	O
application	O
of	O
Soar	B-Architecture
was	O
R1-Soar	O
,	O
a	O
partial	O
reimplementation	O
by	O
Paul	O
Rosenbloom	O
of	O
the	O
R1	B-Application
(	O
XCON	B-Application
)	O
expert	B-Application
system	I-Application
John	O
McDermott	O
developed	O
for	O
configuring	O
DEC	O
computers	O
.	O
</s>
<s>
R1-Soar	O
demonstrated	O
the	O
ability	O
of	O
Soar	B-Architecture
to	O
scale	O
to	O
moderate-size	O
problems	O
,	O
use	O
hierarchical	B-Application
task	I-Application
decomposition	I-Application
and	O
planning	O
,	O
and	O
convert	O
deliberate	O
planning	O
and	O
problem	O
solving	O
to	O
reactive	O
execution	O
through	O
chunking	O
.	O
</s>
<s>
NL-Soar	O
was	O
a	O
natural-language	B-General_Concept
understanding	I-General_Concept
system	O
developed	O
in	O
Soar	B-Architecture
by	O
Jill	O
Fain	O
Lehman	O
,	O
Rick	O
Lewis	O
,	O
Nancy	O
Green	O
,	O
Deryle	O
Lonsdale	O
and	O
Greg	O
Nelson	O
.	O
</s>
<s>
NL-Soar	O
was	O
used	O
in	O
an	O
experimental	O
version	O
of	O
TacAir-Soar	O
and	O
in	O
NTD-Soar	O
.	O
</s>
<s>
The	O
second	O
large-scale	O
application	O
of	O
Soar	B-Architecture
involved	O
developing	O
agents	O
for	O
use	O
in	O
training	O
in	O
large-scale	O
distributed	O
simulation	O
.	O
</s>
<s>
The	O
Michigan	O
system	O
was	O
called	O
TacAir-Soar	O
and	O
flew	O
(	O
in	O
simulation	O
)	O
fixed-wing	O
U	O
.	O
S	O
.	O
military	O
tactical	O
missions	O
(	O
such	O
as	O
close-air	O
support	O
,	O
strikes	O
,	O
CAPs	O
,	O
refueling	O
,	O
and	O
SEAD	O
missions	O
)	O
.	O
</s>
<s>
The	O
ISI	O
system	O
was	O
called	O
RWA-Soar	O
and	O
flew	O
rotary-wing	O
(	O
helicopter	O
)	O
missions	O
.	O
</s>
<s>
Some	O
of	O
the	O
capabilities	O
incorporated	O
in	O
TacAir-Soar	O
and	O
RWA-Soar	O
were	O
attention	O
,	O
situational	O
awareness	O
and	O
adaptation	O
,	O
real-time	O
planning	O
and	O
dynamic	O
replanning	O
,	O
and	O
complex	O
communication	O
,	O
coordination	O
,	O
and	O
cooperation	O
among	O
combinations	O
of	O
Soar	B-Architecture
agents	O
and	O
humans	O
.	O
</s>
<s>
One	O
of	O
the	O
important	O
outgrowths	O
of	O
the	O
RWA-Soar	O
project	O
was	O
the	O
development	O
of	O
STEAM	O
by	O
Milind	O
Tambe	O
,	O
a	O
framework	O
for	O
flexible	O
teamwork	O
in	O
which	O
agents	O
maintained	O
models	O
of	O
their	O
teammates	O
using	O
the	O
joint	O
intentions	O
framework	O
by	O
Cohen	O
&	O
Levesque	O
.	O
</s>
<s>
NTD-Soar	O
was	O
a	O
simulation	O
of	O
the	O
NASA	O
Test	O
Director	O
(	O
NTD	O
)	O
,	O
the	O
person	O
responsible	O
for	O
coordinating	O
the	O
preparation	O
of	O
the	O
NASA	O
Space	O
Shuttle	O
before	O
launch	O
.	O
</s>
<s>
It	O
was	O
an	O
integrated	O
cognitive	O
model	O
that	O
incorporated	O
many	O
different	O
complex	O
cognitive	O
capabilities	O
including	O
natural-language	B-Language
processing	I-Language
,	O
attention	O
and	O
visual	O
search	O
,	O
and	O
problem	O
solving	O
in	O
a	O
broad	O
agent	O
model	O
.	O
</s>
<s>
Soar	B-Architecture
has	O
been	O
used	O
to	O
simulate	O
virtual	O
humans	O
supporting	O
face-to-face	O
dialogues	O
and	O
collaboration	O
within	O
a	O
virtual	O
world	O
developed	O
at	O
the	O
Institute	O
of	O
Creative	O
Technology	O
at	O
USC	O
.	O
</s>
<s>
Virtual	O
humans	O
have	O
integrated	O
capabilities	O
of	O
perception	O
,	O
natural-language	B-General_Concept
understanding	I-General_Concept
,	O
emotions	O
,	O
body	O
control	O
,	O
and	O
action	O
,	O
among	O
others	O
.	O
</s>
<s>
Game	O
AI	O
agents	O
have	O
been	O
built	O
using	O
Soar	B-Architecture
for	O
games	O
such	O
as	O
StarCraft	B-Application
,	O
Quake	B-Application
II	I-Application
,	O
Descent	B-Application
3	I-Application
,	O
Unreal	B-Operating_System
Tournament	I-Operating_System
,	O
and	O
Minecraft	B-Application
,	O
supporting	O
capabilities	O
such	O
as	O
spatial	O
reasoning	O
,	O
real-time	O
strategy	O
,	O
and	O
opponent	O
anticipation	B-General_Concept
.	O
</s>
<s>
AI	O
agents	O
have	O
also	O
been	O
created	O
for	O
video	O
games	O
including	O
Infinite	O
Mario	B-Application
which	O
used	O
reinforcement	O
learning	O
,	O
and	O
Frogger	B-Application
II	I-Application
,	O
Space	B-Application
Invaders	I-Application
,	O
and	O
Fast	O
Eddie	O
,	O
which	O
used	O
both	O
reinforcement	O
learning	O
and	O
mental	O
imagery	O
.	O
</s>
<s>
Soar	B-Architecture
can	O
run	O
natively	O
on	O
mobile	B-Application
devices	I-Application
.	O
</s>
<s>
A	O
mobile	O
for	O
the	O
game	O
Liar	O
’s	O
Dice	O
has	O
been	O
developed	O
for	O
iOS	B-Application
which	O
runs	O
the	O
Soar	B-Architecture
architecture	O
directly	O
from	O
the	O
phone	O
as	O
the	O
engine	O
for	O
opponent	O
AIs	O
.	O
</s>
<s>
Many	O
different	O
robotic	O
applications	O
have	O
been	O
built	O
using	O
Soar	B-Architecture
since	O
the	O
original	O
Robo-Soar	O
was	O
implemented	O
in	O
1991	O
for	O
controlling	O
a	O
Puma	O
robot	O
arm	O
.	O
</s>
<s>
A	O
current	O
focus	O
of	O
research	O
and	O
development	O
in	O
the	O
Soar	B-Architecture
community	O
is	O
Interactive	O
Task	O
Learning	O
(	O
ITL	O
)	O
,	O
the	O
automatic	O
learning	O
of	O
new	O
tasks	O
,	O
environment	O
features	O
,	O
behavioral	O
constraints	O
,	O
and	O
other	O
specifications	O
through	O
natural	O
instructor	O
interaction	O
.	O
</s>
<s>
Early	O
on	O
,	O
Merle-Soar	O
demonstrated	O
how	O
Soar	B-Architecture
could	O
learn	O
a	O
complex	O
scheduling	O
task	O
modeled	O
after	O
the	O
lead	O
human	O
scheduler	O
in	O
a	O
windshield	O
production	O
plant	O
located	O
near	O
Pittsburgh	O
.	O
</s>
<s>
Later	O
,	O
a	O
generalized	O
version	O
of	O
Merle-Soar	O
(	O
Dispatcher-Soar	O
)	O
was	O
used	O
to	O
demonstrate	O
a	O
symbolic	O
,	O
constraint	O
propagation	O
approach	O
in	O
learning	O
to	O
improve	O
schedules	O
and	O
to	O
define	O
task-independent	O
knowledge	O
metrics	O
of	O
architecture-specific	O
learning	O
--	O
knowledge	O
efficiency	O
,	O
knowledge	O
utility	O
,	O
and	O
knowledge	O
effectiveness	O
.	O
</s>
<s>
Melody-Soar	O
demonstrated	O
how	O
the	O
Soar	B-Architecture
architecture	O
could	O
explain	O
and	O
demonstrate	O
creativity	O
in	O
simple	O
melody	O
generation	O
using	O
hierarchies	O
of	O
problems	O
spaces	O
that	O
parallel	O
the	O
hierarchical	O
structure	O
of	O
melody	O
,	O
allowing	O
unique	O
melodies	O
to	O
be	O
generated	O
from	O
preferences	O
of	O
existing	O
styles	O
(	O
e.g.	O
,	O
Bach	O
)	O
.	O
</s>
