<s>
Learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
,	O
or	O
LCS	O
,	O
are	O
a	O
paradigm	O
of	O
rule-based	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
methods	O
that	O
combine	O
a	O
discovery	O
component	O
(	O
e.g.	O
</s>
<s>
typically	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
)	O
with	O
a	O
learning	O
component	O
(	O
performing	O
either	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
reinforcement	O
learning	O
,	O
or	O
unsupervised	B-General_Concept
learning	I-General_Concept
)	O
.	O
</s>
<s>
Learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
seek	O
to	O
identify	O
a	O
set	O
of	O
context-dependent	O
rules	O
that	O
collectively	O
store	O
and	O
apply	O
knowledge	O
in	O
a	O
piecewise	B-Algorithm
manner	O
in	O
order	O
to	O
make	O
predictions	O
(	O
e.g.	O
</s>
<s>
behavior	B-Algorithm
modeling	I-Algorithm
,	O
classification	B-General_Concept
,	O
data	B-Application
mining	I-Application
,	O
regression	O
,	O
function	O
approximation	O
,	O
or	O
game	O
strategy	O
)	O
.	O
</s>
<s>
The	O
founding	O
concepts	O
behind	O
learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
came	O
from	O
attempts	O
to	O
model	O
complex	O
adaptive	O
systems	O
,	O
using	O
rule-based	O
agents	O
to	O
form	O
an	O
artificial	O
cognitive	B-Application
system	I-Application
(	O
i.e.	O
</s>
<s>
artificial	B-Application
intelligence	I-Application
)	O
.	O
</s>
<s>
The	O
architecture	O
and	O
components	O
of	O
a	O
given	O
learning	B-Algorithm
classifier	I-Algorithm
system	I-Algorithm
can	O
be	O
quite	O
variable	O
.	O
</s>
<s>
The	O
major	O
divisions	O
among	O
LCS	O
implementations	O
are	O
as	O
follows	O
:	O
(	O
1	O
)	O
Michigan-style	O
architecture	O
vs.	O
Pittsburgh-style	O
architecture	O
,	O
(	O
2	O
)	O
reinforcement	O
learning	O
vs.	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
(	O
3	O
)	O
incremental	O
learning	O
vs.	O
batch	O
learning	O
,	O
(	O
4	O
)	O
online	B-Algorithm
learning	I-Algorithm
vs.	O
offline	B-General_Concept
learning	I-General_Concept
,	O
(	O
5	O
)	O
strength-based	O
fitness	O
vs.	O
accuracy-based	O
fitness	O
,	O
and	O
(	O
6	O
)	O
complete	O
action	O
mapping	O
vs	O
best	O
action	O
mapping	O
.	O
</s>
<s>
For	O
example	O
,	O
XCS	O
,	O
the	O
best	O
known	O
and	O
best	O
studied	O
LCS	O
algorithm	O
,	O
is	O
Michigan-style	O
,	O
was	O
designed	O
for	O
reinforcement	O
learning	O
but	O
can	O
also	O
perform	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
applies	O
incremental	O
learning	O
that	O
can	O
be	O
either	O
online	O
or	O
offline	O
,	O
applies	O
accuracy-based	O
fitness	O
,	O
and	O
seeks	O
to	O
generate	O
a	O
complete	O
action	O
mapping	O
.	O
</s>
<s>
For	O
simplicity	O
let	O
us	O
focus	O
on	O
Michigan-style	O
architecture	O
with	O
supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
It	O
can	O
be	O
an	O
offline	O
,	O
finite	O
training	O
dataset	O
(	O
characteristic	O
of	O
a	O
data	B-Application
mining	I-Application
,	O
classification	B-General_Concept
,	O
or	O
regression	O
problem	O
)	O
,	O
or	O
an	O
online	O
sequential	O
stream	O
of	O
live	O
training	O
instances	O
.	O
</s>
<s>
Part	O
of	O
LCS	O
learning	O
can	O
involve	O
feature	B-General_Concept
selection	I-General_Concept
,	O
therefore	O
not	O
all	O
of	O
the	O
features	O
in	O
the	O
training	O
data	O
need	O
be	O
informative	O
.	O
</s>
<s>
A	O
critical	O
concept	O
in	O
LCS	O
and	O
rule-based	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
alike	O
,	O
is	O
that	O
an	O
individual	O
rule	O
is	O
not	O
in	O
itself	O
a	O
model	O
,	O
since	O
the	O
rule	O
is	O
only	O
applicable	O
when	O
its	O
condition	O
is	O
satisfied	O
.	O
</s>
<s>
A	O
rule	O
along	O
with	O
its	O
parameters	O
is	O
often	O
referred	O
to	O
as	O
a	O
classifier	B-General_Concept
.	O
</s>
<s>
In	O
Michigan-style	O
systems	O
,	O
classifiers	B-General_Concept
are	O
contained	O
within	O
a	O
population	O
 [ P ] 	O
that	O
has	O
a	O
user	O
defined	O
maximum	O
number	O
of	O
classifiers	B-General_Concept
.	O
</s>
<s>
evolutionary	B-Algorithm
algorithms	I-Algorithm
)	O
,	O
LCS	O
populations	O
start	O
out	O
empty	O
(	O
i.e.	O
</s>
<s>
Classifiers	B-General_Concept
will	O
instead	O
be	O
initially	O
introduced	O
to	O
the	O
population	O
with	O
a	O
covering	O
mechanism	O
.	O
</s>
<s>
In	O
any	O
LCS	O
,	O
the	O
trained	O
model	O
is	O
a	O
set	O
of	O
rules/classifiers	O
,	O
rather	O
than	O
any	O
single	O
rule/classifier	O
.	O
</s>
<s>
In	O
Michigan-style	O
LCS	O
,	O
the	O
entire	O
trained	O
(	O
and	O
optionally	O
,	O
compacted	O
)	O
classifier	B-General_Concept
population	O
forms	O
the	O
prediction	O
model	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
the	O
match	O
set	O
may	O
contain	O
classifiers	B-General_Concept
that	O
propose	O
conflicting	O
actions	O
.	O
</s>
<s>
In	O
the	O
fourth	O
step	O
,	O
since	O
we	O
are	O
performing	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
 [ M ] 	O
is	O
divided	O
into	O
a	O
correct	O
set	O
 [ C ] 	O
and	O
an	O
incorrect	O
set	O
 [ I ] 	O
.	O
</s>
<s>
At	O
this	O
point	O
in	O
the	O
learning	O
cycle	O
,	O
if	O
no	O
classifiers	B-General_Concept
made	O
it	O
into	O
either	O
 [ M ] 	O
or	O
 [ C ] 	O
(	O
as	O
would	O
be	O
the	O
case	O
when	O
the	O
population	O
starts	O
off	O
empty	O
)	O
,	O
the	O
covering	O
mechanism	O
is	O
applied	O
(	O
fifth	O
step	O
)	O
.	O
</s>
<s>
Covering	O
randomly	O
generates	O
a	O
rule	O
that	O
matches	O
the	O
current	O
training	O
instance	O
(	O
and	O
in	O
the	O
case	O
of	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
that	O
rule	O
is	O
also	O
generated	O
with	O
the	O
correct	O
action	O
.	O
</s>
<s>
For	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
we	O
can	O
simply	O
update	O
the	O
accuracy/error	O
of	O
a	O
rule	O
.	O
</s>
<s>
The	O
concept	O
of	O
fitness	O
is	O
taken	O
directly	O
from	O
classic	O
genetic	B-Algorithm
algorithms	I-Algorithm
.	O
</s>
<s>
Subsumption	O
is	O
an	O
explicit	O
generalization	O
mechanism	O
that	O
merges	O
classifiers	B-General_Concept
that	O
cover	O
redundant	O
parts	O
of	O
the	O
problem	O
space	O
.	O
</s>
<s>
The	O
subsuming	O
classifier	B-General_Concept
effectively	O
absorbs	O
the	O
subsumed	O
classifier	B-General_Concept
(	O
and	O
has	O
its	O
numerosity	O
increased	O
)	O
.	O
</s>
<s>
This	O
can	O
only	O
happen	O
when	O
the	O
subsuming	O
classifier	B-General_Concept
is	O
more	O
general	O
,	O
just	O
as	O
accurate	O
,	O
and	O
covers	O
all	O
of	O
the	O
problem	O
space	O
of	O
the	O
classifier	B-General_Concept
it	O
subsumes	O
.	O
</s>
<s>
In	O
the	O
eighth	O
step	O
,	O
LCS	O
adopts	O
a	O
highly	O
elitist	O
genetic	B-Algorithm
algorithm	I-Algorithm
(	O
GA	O
)	O
which	O
will	O
select	O
two	O
parent	O
classifiers	B-General_Concept
based	O
on	O
fitness	O
(	O
survival	O
of	O
the	O
fittest	O
)	O
.	O
</s>
<s>
The	O
LCS	O
genetic	B-Algorithm
algorithm	I-Algorithm
is	O
highly	O
elitist	O
since	O
each	O
learning	O
iteration	O
,	O
the	O
vast	O
majority	O
of	O
the	O
population	O
is	O
preserved	O
.	O
</s>
<s>
Rule	B-Algorithm
discovery	I-Algorithm
may	O
alternatively	O
be	O
performed	O
by	O
some	O
other	O
method	O
,	O
such	O
as	O
an	O
estimation	O
of	O
distribution	O
algorithm	O
,	O
but	O
a	O
GA	O
is	O
by	O
far	O
the	O
most	O
common	O
approach	O
.	O
</s>
<s>
Evolutionary	B-Algorithm
algorithms	I-Algorithm
like	O
the	O
GA	O
employ	O
a	O
stochastic	O
search	O
,	O
which	O
makes	O
LCS	O
a	O
stochastic	O
algorithm	O
.	O
</s>
<s>
The	O
deletion	O
mechanism	O
will	O
select	O
classifiers	B-General_Concept
for	O
deletion	O
(	O
commonly	O
using	O
roulette	O
wheel	O
selection	O
)	O
.	O
</s>
<s>
The	O
probability	O
of	O
a	O
classifier	B-General_Concept
being	O
selected	O
for	O
deletion	O
is	O
inversely	O
proportional	O
to	O
its	O
fitness	O
.	O
</s>
<s>
When	O
a	O
classifier	B-General_Concept
is	O
selected	O
for	O
deletion	O
,	O
its	O
numerosity	O
parameter	O
is	O
reduced	O
by	O
one	O
.	O
</s>
<s>
When	O
the	O
numerosity	O
of	O
a	O
classifier	B-General_Concept
is	O
reduced	O
to	O
zero	O
,	O
it	O
is	O
removed	O
entirely	O
from	O
the	O
population	O
.	O
</s>
<s>
For	O
online	B-Algorithm
learning	I-Algorithm
,	O
LCS	O
will	O
obtain	O
a	O
completely	O
new	O
training	O
instance	O
each	O
iteration	O
from	O
the	O
environment	O
.	O
</s>
<s>
For	O
offline	B-General_Concept
learning	I-General_Concept
,	O
LCS	O
will	O
iterate	O
through	O
a	O
finite	O
training	O
dataset	O
.	O
</s>
<s>
Whether	O
or	O
not	O
rule	O
compaction	O
has	O
been	O
applied	O
,	O
the	O
output	O
of	O
an	O
LCS	O
algorithm	O
is	O
a	O
population	O
of	O
classifiers	B-General_Concept
which	O
can	O
be	O
applied	O
to	O
making	O
predictions	O
on	O
previously	O
unseen	O
instances	O
.	O
</s>
<s>
With	O
respect	O
to	O
other	O
advanced	O
machine	O
learning	O
approaches	O
,	O
such	O
as	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
random	B-Algorithm
forests	I-Algorithm
,	O
or	O
genetic	B-Algorithm
programming	I-Algorithm
,	O
learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
are	O
particularly	O
well	O
suited	O
to	O
problems	O
that	O
require	O
interpretable	O
solutions	O
.	O
</s>
<s>
John	O
Henry	O
Holland	O
was	O
best	O
known	O
for	O
his	O
work	O
popularizing	O
genetic	B-Algorithm
algorithms	I-Algorithm
(	O
GA	O
)	O
,	O
through	O
his	O
ground-breaking	O
book	O
"	O
Adaptation	O
in	O
Natural	O
and	O
Artificial	O
Systems	O
"	O
in	O
1975	O
and	O
his	O
formalization	O
of	O
Holland	O
's	O
schema	O
theorem	O
.	O
</s>
<s>
In	O
1976	O
,	O
Holland	O
conceptualized	O
an	O
extension	O
of	O
the	O
GA	O
concept	O
to	O
what	O
he	O
called	O
a	O
"	O
cognitive	B-Application
system	I-Application
"	O
,	O
and	O
provided	O
the	O
first	O
detailed	O
description	O
of	O
what	O
would	O
become	O
known	O
as	O
the	O
first	O
learning	B-Algorithm
classifier	I-Algorithm
system	I-Algorithm
in	O
the	O
paper	O
"	O
Cognitive	B-Application
Systems	I-Application
based	O
on	O
Adaptive	O
Algorithms	O
"	O
.	O
</s>
<s>
This	O
first	O
system	O
,	O
named	O
Cognitive	B-Application
System	I-Application
One	O
(	O
CS-1	O
)	O
was	O
conceived	O
as	O
a	O
modeling	O
tool	O
,	O
designed	O
to	O
model	O
a	O
real	O
system	O
(	O
i.e.	O
</s>
<s>
The	O
goal	O
was	O
for	O
a	O
set	O
of	O
rules	O
to	O
perform	O
online	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
to	O
adapt	O
to	O
the	O
environment	O
based	O
on	O
infrequent	O
payoff/reward	O
(	O
i.e.	O
</s>
<s>
Beginning	O
in	O
1980	O
,	O
Kenneth	O
de	O
Jong	O
and	O
his	O
student	O
Stephen	O
Smith	O
took	O
a	O
different	O
approach	O
to	O
rule-based	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
with	O
(	O
LS-1	O
)	O
,	O
where	O
learning	O
was	O
viewed	O
as	O
an	O
offline	O
optimization	O
process	O
rather	O
than	O
an	O
online	O
adaptation	O
process	O
.	O
</s>
<s>
This	O
new	O
approach	O
was	O
more	O
similar	O
to	O
a	O
standard	O
genetic	B-Algorithm
algorithm	I-Algorithm
but	O
evolved	O
independent	O
sets	O
of	O
rules	O
.	O
</s>
<s>
Since	O
that	O
time	O
LCS	O
methods	O
inspired	O
by	O
the	O
online	B-Algorithm
learning	I-Algorithm
framework	O
introduced	O
by	O
Holland	O
at	O
the	O
University	O
of	O
Michigan	O
have	O
been	O
referred	O
to	O
as	O
Michigan-style	O
LCS	O
,	O
and	O
those	O
inspired	O
by	O
Smith	O
and	O
De	O
Jong	O
at	O
the	O
University	O
of	O
Pittsburgh	O
have	O
been	O
referred	O
to	O
as	O
Pittsburgh-style	O
LCS	O
.	O
</s>
<s>
the	O
match	O
set	O
 [ M ] 	O
)	O
rather	O
than	O
from	O
the	O
whole	O
population	O
[P],	O
(	O
3	O
)	O
covering	O
,	O
first	O
introduced	O
as	O
a	O
create	O
operator	O
,	O
(	O
4	O
)	O
the	O
formalization	O
of	O
an	O
action	O
set	O
[A],	O
(	O
5	O
)	O
a	O
simplified	O
algorithm	O
architecture	O
,	O
(	O
6	O
)	O
strength-based	O
fitness	O
,	O
(	O
7	O
)	O
consideration	O
of	O
single-step	O
,	O
or	O
supervised	B-General_Concept
learning	I-General_Concept
problems	O
and	O
the	O
introduction	O
of	O
the	O
correct	O
set	O
[C],	O
(	O
8	O
)	O
accuracy-based	O
fitness	O
(	O
9	O
)	O
the	O
combination	O
of	O
fuzzy	O
logic	O
with	O
LCS	O
(	O
which	O
later	O
spawned	O
a	O
lineage	O
of	O
fuzzy	O
LCS	O
algorithms	O
)	O
,	O
(	O
10	O
)	O
encouraging	O
long	O
action	O
chains	O
and	O
default	O
hierarchies	O
for	O
improving	O
performance	O
on	O
multi-step	O
problems	O
,	O
(	O
11	O
)	O
examining	O
latent	O
learning	O
(	O
which	O
later	O
inspired	O
a	O
new	O
branch	O
of	O
anticipatory	O
classifier	B-Algorithm
systems	I-Algorithm
(	O
ACS	O
)	O
)	O
,	O
and	O
(	O
12	O
)	O
the	O
introduction	O
of	O
the	O
first	O
Q-learning-like	O
credit	O
assignment	O
technique	O
.	O
</s>
<s>
Interest	O
in	O
learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
was	O
reinvigorated	O
in	O
the	O
mid	O
1990s	O
largely	O
due	O
to	O
two	O
events	O
;	O
the	O
development	O
of	O
the	O
Q-Learning	B-Algorithm
algorithm	O
for	O
reinforcement	O
learning	O
,	O
and	O
the	O
introduction	O
of	O
significantly	O
simplified	O
Michigan-style	O
LCS	O
architectures	O
by	O
Stewart	O
Wilson	O
.	O
</s>
<s>
Wilson	O
's	O
Zeroth-level	O
Classifier	B-Algorithm
System	I-Algorithm
(	O
ZCS	O
)	O
focused	O
on	O
increasing	O
algorithmic	O
understandability	O
based	O
on	O
Hollands	O
standard	O
LCS	O
implementation	O
.	O
</s>
<s>
This	O
was	O
done	O
,	O
in	O
part	O
,	O
by	O
removing	O
rule-bidding	O
and	O
the	O
internal	O
message	O
list	O
,	O
essential	O
to	O
the	O
original	O
BBA	O
credit	O
assignment	O
,	O
and	O
replacing	O
it	O
with	O
a	O
hybrid	O
BBA/Q	O
-Learning	O
strategy	O
.	O
</s>
<s>
However	O
,	O
ZCS	O
still	O
suffered	O
from	O
performance	O
drawbacks	O
including	O
the	O
proliferation	O
of	O
over-general	O
classifiers	B-General_Concept
.	O
</s>
<s>
In	O
1995	O
,	O
Wilson	O
published	O
his	O
landmark	O
paper	O
,	O
"	O
Classifier	B-General_Concept
fitness	O
based	O
on	O
accuracy	O
"	O
in	O
which	O
he	O
introduced	O
the	O
classifier	B-Algorithm
system	I-Algorithm
XCS	O
.	O
</s>
<s>
XCS	O
took	O
the	O
simplified	O
architecture	O
of	O
ZCS	O
and	O
added	O
an	O
accuracy-based	O
fitness	O
,	O
a	O
niche	O
GA	O
(	O
acting	O
in	O
the	O
action	O
set	O
 [ A ] 	O
)	O
,	O
an	O
explicit	O
generalization	O
mechanism	O
called	O
subsumption	O
,	O
and	O
an	O
adaptation	O
of	O
the	O
Q-Learning	B-Algorithm
credit	O
assignment	O
.	O
</s>
<s>
XCS	O
was	O
popularized	O
by	O
its	O
ability	O
to	O
reach	O
optimal	O
performance	O
while	O
evolving	O
accurate	O
and	O
maximally	O
general	O
classifiers	B-General_Concept
as	O
well	O
as	O
its	O
impressive	O
problem	O
flexibility	O
(	O
able	O
to	O
perform	O
both	O
reinforcement	O
learning	O
and	O
supervised	B-General_Concept
learning	I-General_Concept
)	O
.	O
</s>
<s>
Differently	O
,	O
most	O
strength-based	O
LCSs	O
,	O
or	O
exclusively	O
supervised	B-General_Concept
learning	I-General_Concept
LCSs	O
seek	O
a	O
rule	O
set	O
of	O
efficient	O
generalizations	O
in	O
the	O
form	O
of	O
a	O
best	O
action	O
map	O
(	O
or	O
a	O
partial	O
map	O
)	O
.	O
</s>
<s>
In	O
1995	O
,	O
Congdon	O
was	O
the	O
first	O
to	O
apply	O
LCS	O
to	O
real-world	O
epidemiological	O
investigations	O
of	O
disease	O
followed	O
closely	O
by	O
Holmes	O
who	O
developed	O
the	O
BOOLE++	O
,	O
EpiCS	O
,	O
and	O
later	O
EpiXCS	O
for	O
epidemiological	O
classification	B-General_Concept
.	O
</s>
<s>
These	O
early	O
works	O
inspired	O
later	O
interest	O
in	O
applying	O
LCS	O
algorithms	O
to	O
complex	O
and	O
large-scale	O
data	B-Application
mining	I-Application
tasks	O
epitomized	O
by	O
bioinformatics	O
applications	O
.	O
</s>
<s>
In	O
1998	O
,	O
Stolzmann	O
introduced	O
anticipatory	O
classifier	B-Algorithm
systems	I-Algorithm
(	O
ACS	O
)	O
which	O
included	O
rules	O
in	O
the	O
form	O
of	O
'	O
condition-action-effect	O
,	O
rather	O
than	O
the	O
classic	O
'	O
condition-action	O
'	O
representation	O
.	O
</s>
<s>
In	O
2003	O
,	O
Bernado-Mansilla	O
introduced	O
a	O
sUpervised	O
Classifier	B-Algorithm
System	I-Algorithm
(	O
UCS	O
)	O
,	O
which	O
specialized	O
the	O
XCS	O
algorithm	O
to	O
the	O
task	O
of	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
single-step	O
problems	O
,	O
and	O
forming	O
a	O
best	O
action	O
set	O
.	O
</s>
<s>
Bull	O
introduced	O
a	O
simple	O
accuracy-based	O
LCS	O
(	O
YCS	O
)	O
and	O
a	O
simple	O
strength-based	O
LCS	O
Minimal	O
Classifier	B-Algorithm
System	I-Algorithm
(	O
MCS	O
)	O
in	O
order	O
to	O
develop	O
a	O
better	O
theoretical	O
understanding	O
of	O
the	O
LCS	O
framework	O
.	O
</s>
<s>
Bacardit	O
introduced	O
GAssist	O
and	O
BioHEL	O
,	O
Pittsburgh-style	O
LCSs	O
designed	O
for	O
data	B-Application
mining	I-Application
and	O
scalability	B-Architecture
to	O
large	O
datasets	O
in	O
bioinformatics	O
applications	O
.	O
</s>
<s>
In	O
2008	O
,	O
Drugowitsch	O
published	O
the	O
book	O
titled	O
"	O
Design	O
and	O
Analysis	O
of	O
Learning	B-Algorithm
Classifier	I-Algorithm
Systems	I-Algorithm
"	O
including	O
some	O
theoretical	O
examination	O
of	O
LCS	O
algorithms	O
.	O
</s>
<s>
Butz	O
introduced	O
the	O
first	O
rule	O
online	B-Algorithm
learning	I-Algorithm
visualization	O
within	O
a	O
GUI	B-Application
for	O
XCSF	O
(	O
see	O
the	O
image	O
at	O
the	O
top	O
of	O
this	O
page	O
)	O
.	O
</s>
<s>
Urbanowicz	O
extended	O
the	O
UCS	O
framework	O
and	O
introduced	O
ExSTraCS	O
,	O
explicitly	O
designed	O
for	O
supervised	B-General_Concept
learning	I-General_Concept
in	O
noisy	O
problem	O
domains	O
(	O
e.g.	O
</s>
<s>
ExSTraCS	O
integrated	O
(	O
1	O
)	O
expert	O
knowledge	O
to	O
drive	O
covering	O
and	O
genetic	B-Algorithm
algorithm	I-Algorithm
towards	O
important	O
features	O
in	O
the	O
data	O
,	O
(	O
2	O
)	O
a	O
form	O
of	O
long-term	O
memory	O
referred	O
to	O
as	O
attribute	O
tracking	O
,	O
allowing	O
for	O
more	O
efficient	O
learning	O
and	O
the	O
characterization	O
of	O
heterogeneous	B-General_Concept
data	O
patterns	O
,	O
and	O
(	O
3	O
)	O
a	O
flexible	O
rule	O
representation	O
similar	O
to	O
Bacardit	O
's	O
mixed	O
discrete-continuous	O
attribute	O
list	O
representation	O
.	O
</s>
<s>
Both	O
Bacardit	O
and	O
Urbanowicz	O
explored	O
statistical	O
and	O
visualization	O
strategies	O
to	O
interpret	O
LCS	O
rules	O
and	O
perform	O
knowledge	O
discovery	O
for	O
data	B-Application
mining	I-Application
.	O
</s>
<s>
Browne	O
and	O
Iqbal	O
explored	O
the	O
concept	O
of	O
reusing	O
building	O
blocks	O
in	O
the	O
form	O
of	O
code	O
fragments	O
and	O
were	O
the	O
first	O
to	O
solve	O
the	O
135-bit	O
multiplexer	B-Protocol
benchmark	O
problem	O
by	O
first	O
learning	O
useful	O
building	O
blocks	O
from	O
simpler	O
multiplexer	B-Protocol
problems	O
.	O
</s>
<s>
ExSTraCS	O
2.0	O
was	O
later	O
introduced	O
to	O
improve	O
Michigan-style	O
LCS	O
scalability	B-Architecture
,	O
successfully	O
solving	O
the	O
135-bit	O
multiplexer	B-Protocol
benchmark	O
problem	O
for	O
the	O
first	O
time	O
directly	O
.	O
</s>
<s>
The	O
n-bit	O
multiplexer	B-Protocol
problem	O
is	O
highly	O
epistatic	O
and	O
heterogeneous	B-General_Concept
,	O
making	O
it	O
a	O
very	O
challenging	O
machine	O
learning	O
task	O
.	O
</s>
<s>
Michigan-Style	O
LCSs	O
are	O
characterized	O
by	O
a	O
population	O
of	O
rules	O
where	O
the	O
genetic	B-Algorithm
algorithm	I-Algorithm
operates	O
at	O
the	O
level	O
of	O
individual	O
rules	O
and	O
the	O
solution	O
is	O
represented	O
by	O
the	O
entire	O
rule	O
population	O
.	O
</s>
<s>
Michigan	O
style	O
systems	O
also	O
learn	O
incrementally	O
which	O
allows	O
them	O
to	O
perform	O
both	O
reinforcement	O
learning	O
and	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
as	O
well	O
as	O
both	O
online	O
and	O
offline	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
The	O
genetic	B-Algorithm
algorithm	I-Algorithm
typically	O
operates	O
at	O
the	O
level	O
of	O
an	O
entire	O
rule-set	O
.	O
</s>
<s>
Adaptive	O
:	O
They	O
can	O
acclimate	O
to	O
a	O
changing	O
environment	O
in	O
the	O
case	O
of	O
online	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
They	O
can	O
model	O
complex	O
,	O
epistatic	O
,	O
heterogeneous	B-General_Concept
,	O
or	O
distributed	O
underlying	O
patterns	O
without	O
relying	O
on	O
prior	O
knowledge	O
.	O
</s>
<s>
Interpretable:In	O
the	O
interest	O
of	O
data	B-Application
mining	I-Application
and	O
knowledge	O
discovery	O
individual	O
LCS	O
rules	O
are	O
logical	O
,	O
and	O
can	O
be	O
made	O
to	O
be	O
human	O
interpretable	O
IF:THEN	O
statements	O
.	O
</s>
<s>
Overfitting	B-Error_Name
:	O
Like	O
any	O
machine	O
learner	O
,	O
LCS	O
can	O
suffer	O
from	O
overfitting	B-Error_Name
despite	O
implicit	O
and	O
explicit	O
generalization	O
pressures	O
.	O
</s>
<s>
The	O
name	O
,	O
"	O
Learning	B-Algorithm
Classifier	I-Algorithm
System	I-Algorithm
(	O
LCS	O
)	O
"	O
,	O
is	O
a	O
bit	O
misleading	O
since	O
there	O
are	O
many	O
machine	O
learning	O
algorithms	O
that	O
'	O
learn	O
to	O
classify	O
 '	O
(	O
e.g.	O
</s>
<s>
decision	B-Algorithm
trees	I-Algorithm
,	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
)	O
,	O
but	O
are	O
not	O
LCSs	O
.	O
</s>
<s>
The	O
term	O
'	O
rule-based	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
(	O
RBML	B-Algorithm
)	O
'	O
is	O
useful	O
,	O
as	O
it	O
more	O
clearly	O
captures	O
the	O
essential	O
'	O
rule-based	O
'	O
component	O
of	O
these	O
systems	O
,	O
but	O
it	O
also	O
generalizes	O
to	O
methods	O
that	O
are	O
not	O
considered	O
to	O
be	O
LCSs	O
(	O
e.g.	O
</s>
<s>
association	B-Algorithm
rule	I-Algorithm
learning	I-Algorithm
,	O
or	O
artificial	B-Algorithm
immune	I-Algorithm
systems	I-Algorithm
)	O
.	O
</s>
<s>
More	O
general	O
terms	O
such	O
as	O
,	O
'	O
genetics-based	O
machine	O
learning	O
 '	O
,	O
and	O
even	O
'	O
genetic	B-Algorithm
algorithm	I-Algorithm
 '	O
have	O
also	O
been	O
applied	O
to	O
refer	O
to	O
what	O
would	O
be	O
more	O
characteristically	O
defined	O
as	O
a	O
learning	B-Algorithm
classifier	I-Algorithm
system	I-Algorithm
.	O
</s>
<s>
Due	O
to	O
their	O
similarity	O
to	O
genetic	B-Algorithm
algorithms	I-Algorithm
,	O
Pittsburgh-style	O
learning	B-Algorithm
classifier	I-Algorithm
systems	I-Algorithm
are	O
sometimes	O
generically	O
referred	O
to	O
as	O
'	O
genetic	B-Algorithm
algorithms	I-Algorithm
 '	O
.	O
</s>
<s>
Beyond	O
this	O
,	O
some	O
LCS	O
algorithms	O
,	O
or	O
closely	O
related	O
methods	O
,	O
have	O
been	O
referred	O
to	O
as	O
'	O
cognitive	B-Application
systems	I-Application
 '	O
,	O
'	O
adaptive	O
agents	O
 '	O
,	O
'	O
production	B-Application
systems	I-Application
 '	O
,	O
or	O
generically	O
as	O
a	O
'	O
classifier	B-Algorithm
system	I-Algorithm
 '	O
.	O
</s>
<s>
Up	O
until	O
the	O
2000s	O
nearly	O
all	O
learning	B-Algorithm
classifier	I-Algorithm
system	I-Algorithm
methods	O
were	O
developed	O
with	O
reinforcement	O
learning	O
problems	O
in	O
mind	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
the	O
term	O
‘	O
learning	O
classifier	O
system’	O
was	O
commonly	O
defined	O
as	O
the	O
combination	O
of	O
‘	O
trial-and-error	O
’	O
reinforcement	O
learning	O
with	O
the	O
global	O
search	O
of	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
Interest	O
in	O
supervised	B-General_Concept
learning	I-General_Concept
applications	O
,	O
and	O
even	O
unsupervised	B-General_Concept
learning	I-General_Concept
have	O
since	O
broadened	O
the	O
use	O
and	O
definition	O
of	O
this	O
term	O
.	O
</s>
