<s>
Multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
(	O
MARL	O
)	O
is	O
a	O
sub-field	O
of	O
reinforcement	O
learning	O
.	O
</s>
<s>
Multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
is	O
closely	O
related	O
to	O
game	O
theory	O
and	O
especially	O
repeated	O
games	O
,	O
as	O
well	O
as	O
multi-agent	O
systems	O
.	O
</s>
<s>
While	O
research	O
in	O
single-agent	O
reinforcement	O
learning	O
is	O
concerned	O
with	O
finding	O
the	O
algorithm	O
that	O
gets	O
the	O
biggest	O
number	O
of	O
points	O
for	O
one	O
agent	O
,	O
research	O
in	O
multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
evaluates	O
and	O
quantifies	O
social	O
metrics	O
,	O
such	O
as	O
cooperation	O
,	O
reciprocity	O
,	O
equity	O
,	O
social	O
influence	O
,	O
language	O
and	O
discrimination	O
.	O
</s>
<s>
Similarly	O
to	O
single-agent	O
reinforcement	O
learning	O
,	O
multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
is	O
modeled	O
as	O
some	O
form	O
of	O
a	O
Markov	O
decision	O
process	O
(	O
MDP	O
)	O
.	O
</s>
<s>
In	O
settings	O
with	O
perfect	O
information	O
,	O
such	O
as	O
the	O
games	O
of	O
chess	B-Application
and	O
Go	O
,	O
the	O
MDP	O
would	O
be	O
fully	O
observable	O
.	O
</s>
<s>
Many	O
traditional	O
games	O
such	O
as	O
chess	B-Application
and	O
Go	O
fall	O
under	O
this	O
category	O
,	O
as	O
do	O
two-player	O
variants	O
of	O
modern	O
games	O
like	O
StarCraft	B-Application
.	O
</s>
<s>
The	O
Deep	B-General_Concept
Blue	I-General_Concept
and	O
AlphaGo	B-Application
projects	O
demonstrate	O
how	O
to	O
optimize	O
the	O
performance	O
of	O
agents	O
in	O
pure	O
competition	O
settings	O
.	O
</s>
<s>
As	O
the	O
agents	O
 '	O
policy	O
is	O
improved	O
using	O
self-play	B-Algorithm
,	O
multiple	O
layers	O
of	O
learning	O
may	O
occur	O
.	O
</s>
<s>
Pure	O
cooperation	O
settings	O
are	O
explored	O
in	O
recreational	O
cooperative	O
games	O
such	O
as	O
Overcooked	B-Application
,	O
as	O
well	O
as	O
real-world	O
scenarios	O
in	O
robotics	O
.	O
</s>
<s>
StarCraft	B-Application
II	I-Application
.	O
</s>
<s>
The	O
seekers	O
then	O
respond	O
by	O
"	O
box	O
surfing	O
"	O
,	O
exploiting	O
a	O
glitch	B-Error_Name
in	O
the	O
game	O
to	O
penetrate	O
the	O
shelter	O
.	O
</s>
<s>
Multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
has	O
been	O
applied	O
to	O
a	O
variety	O
of	O
use	O
cases	O
in	O
science	O
and	O
industry	O
:	O
</s>
<s>
Multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
has	O
been	O
used	O
in	O
research	O
into	O
AI	O
alignment	O
.	O
</s>
<s>
There	O
are	O
some	O
inherent	O
difficulties	O
about	O
multi-agent	O
deep	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
There	O
are	O
various	O
tools	O
and	O
frameworks	O
for	O
working	O
with	O
multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
environments	O
:	O
</s>
