<s>
In	O
video	O
games	O
,	O
various	O
artificial	B-Application
intelligence	I-Application
techniques	O
have	O
been	O
used	O
in	O
a	O
variety	O
of	O
ways	O
,	O
ranging	O
from	O
non-player	O
character	O
(	O
NPC	O
)	O
control	O
to	O
procedural	O
content	O
generation	O
(	O
PCG	O
)	O
.	O
</s>
<s>
Machine	O
learning	O
is	O
a	O
subset	O
of	O
artificial	B-Application
intelligence	I-Application
that	O
focuses	O
on	O
using	O
algorithms	O
and	O
statistical	O
models	O
to	O
make	O
machines	O
act	O
without	O
specific	O
programming	O
.	O
</s>
<s>
This	O
is	O
in	O
sharp	O
contrast	O
to	O
traditional	O
methods	O
of	O
artificial	B-Application
intelligence	I-Application
such	O
as	O
search	B-Data_Structure
trees	I-Data_Structure
and	O
expert	B-Application
systems	I-Application
.	O
</s>
<s>
The	O
most	O
publicly	O
known	O
application	O
of	O
machine	O
learning	O
in	O
games	O
is	O
likely	O
the	O
use	O
of	O
deep	B-Algorithm
learning	I-Algorithm
agents	B-General_Concept
that	O
compete	O
with	O
professional	O
human	O
players	O
in	O
complex	O
strategy	B-Application
games	I-Application
.	O
</s>
<s>
There	O
has	O
been	O
a	O
significant	O
application	O
of	O
machine	O
learning	O
on	O
games	O
such	O
as	O
Atari/ALE	O
,	O
Doom	B-Application
,	O
Minecraft	B-Application
,	O
StarCraft	B-Application
,	O
and	O
car	O
racing	O
.	O
</s>
<s>
Other	O
games	O
that	O
did	O
not	O
originally	O
exists	O
as	O
video	O
games	O
,	O
such	O
as	O
chess	B-Application
and	O
Go	O
have	O
also	O
been	O
affected	O
by	O
the	O
machine	O
learning	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
a	O
subset	O
of	O
machine	O
learning	O
which	O
focuses	O
heavily	O
on	O
the	O
use	O
of	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
ANN	O
)	O
that	O
learn	O
to	O
solve	O
complex	O
tasks	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
uses	O
multiple	O
layers	O
of	O
ANN	O
and	O
other	O
techniques	O
to	O
progressively	O
extract	O
information	O
from	O
an	O
input	O
.	O
</s>
<s>
Due	O
to	O
this	O
complex	O
layered	O
approach	O
,	O
deep	B-Algorithm
learning	I-Algorithm
models	O
often	O
require	O
powerful	O
machines	O
to	O
train	O
and	O
run	O
on	O
.	O
</s>
<s>
Convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
CNN	B-Architecture
)	O
are	O
specialized	O
ANNs	O
that	O
are	O
often	O
used	O
to	O
analyze	O
image	O
data	O
.	O
</s>
<s>
CNNs	B-Architecture
are	O
able	O
to	O
learn	O
these	O
patterns	O
in	O
a	O
hierarchy	O
,	O
meaning	O
that	O
earlier	O
convolutional	O
layers	O
will	O
learn	O
smaller	O
local	O
patterns	O
while	O
later	O
layers	O
will	O
learn	O
larger	O
patterns	O
based	O
on	O
the	O
previous	O
patterns	O
.	O
</s>
<s>
A	O
CNN	B-Architecture
's	O
ability	O
to	O
learn	O
visual	O
data	O
has	O
made	O
it	O
a	O
commonly	O
used	O
tool	O
for	O
deep	B-Algorithm
learning	I-Algorithm
in	O
games	O
.	O
</s>
<s>
Recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
are	O
a	O
type	O
of	O
ANN	O
that	O
are	O
designed	O
to	O
process	O
sequences	O
of	O
data	O
in	O
order	O
,	O
one	O
part	O
at	O
a	O
time	O
rather	O
than	O
all	O
at	O
once	O
.	O
</s>
<s>
These	O
types	O
of	O
ANN	O
are	O
highly	O
effective	O
at	O
tasks	O
such	O
as	O
speech	B-Application
recognition	I-Application
and	O
other	O
problems	O
that	O
depend	O
heavily	O
on	O
temporal	O
order	O
.	O
</s>
<s>
There	O
are	O
several	O
types	O
of	O
RNNs	O
with	O
different	O
internal	O
configurations	O
;	O
the	O
basic	O
implementation	O
suffers	O
from	O
a	O
lack	O
of	O
long	O
term	O
memory	O
due	O
to	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
,	O
thus	O
it	O
is	O
rarely	O
used	O
over	O
newer	O
implementations	O
.	O
</s>
<s>
A	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
network	O
is	O
a	O
specific	O
implementation	O
of	O
a	O
RNN	O
that	O
is	O
designed	O
to	O
deal	O
with	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
seen	O
in	O
simple	O
RNNs	O
,	O
which	O
would	O
lead	O
to	O
them	O
gradually	O
"	O
forgetting	O
"	O
about	O
previous	O
parts	O
of	O
an	O
inputted	O
sequence	O
when	O
calculating	O
the	O
output	O
of	O
a	O
current	O
part	O
.	O
</s>
<s>
LSTMs	B-Algorithm
solve	O
this	O
problem	O
with	O
the	O
addition	O
of	O
an	O
elaborate	O
system	O
that	O
uses	O
an	O
additional	O
input/output	O
to	O
keep	O
track	O
of	O
long	O
term	O
data	O
.	O
</s>
<s>
LSTMs	B-Algorithm
have	O
achieved	O
very	O
strong	O
results	O
across	O
various	O
fields	O
,	O
and	O
were	O
used	O
by	O
several	O
monumental	O
deep	B-Algorithm
learning	I-Algorithm
agents	B-General_Concept
in	O
games	O
.	O
</s>
<s>
Reinforcement	O
learning	O
is	O
used	O
heavily	O
in	O
the	O
field	O
of	O
machine	O
learning	O
and	O
can	O
be	O
seen	O
in	O
methods	O
such	O
as	O
Q-learning	B-Algorithm
,	O
policy	O
search	O
,	O
Deep	O
Q-networks	O
and	O
others	O
.	O
</s>
<s>
Neuroevolution	B-Algorithm
involves	O
the	O
use	O
of	O
both	O
neural	B-Architecture
networks	I-Architecture
and	O
evolutionary	B-Algorithm
algorithms	I-Algorithm
.	O
</s>
<s>
Instead	O
of	O
using	O
gradient	O
descent	O
like	O
most	O
neural	B-Architecture
networks	I-Architecture
,	O
neuroevolution	B-Algorithm
models	O
make	O
use	O
of	O
evolutionary	B-Algorithm
algorithms	I-Algorithm
to	O
update	O
neurons	O
in	O
the	O
network	O
.	O
</s>
<s>
Researchers	O
claim	O
that	O
this	O
process	O
is	O
less	O
likely	O
to	O
get	O
stuck	O
in	O
a	O
local	O
minimum	O
and	O
is	O
potentially	O
faster	O
than	O
state	O
of	O
the	O
art	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
.	O
</s>
<s>
Machine	O
learning	O
agents	B-General_Concept
have	O
been	O
used	O
to	O
take	O
the	O
place	O
of	O
a	O
human	O
player	O
rather	O
than	O
function	O
as	O
NPCs	O
,	O
which	O
are	O
deliberately	O
added	O
into	O
video	O
games	O
as	O
part	O
of	O
designed	O
gameplay	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
agents	B-General_Concept
have	O
achieved	O
impressive	O
results	O
when	O
used	O
in	O
competition	O
with	O
both	O
humans	O
and	O
other	O
artificial	B-General_Concept
intelligence	I-General_Concept
agents	I-General_Concept
.	O
</s>
<s>
Chess	B-Application
is	O
a	O
turn-based	O
strategy	B-Application
game	I-Application
that	O
is	O
considered	O
a	O
difficult	O
AI	B-Application
problem	O
due	O
to	O
the	O
computational	O
complexity	O
of	O
its	O
board	O
space	O
.	O
</s>
<s>
Similar	O
strategy	B-Application
games	I-Application
are	O
often	O
solved	O
with	O
some	O
form	O
of	O
a	O
Minimax	B-Algorithm
Tree	O
Search	O
.	O
</s>
<s>
These	O
types	O
of	O
AI	B-Application
agents	B-General_Concept
have	O
been	O
known	O
to	O
beat	O
professional	O
human	O
players	O
,	O
such	O
as	O
the	O
historic	O
1997	O
Deep	O
Blue	O
versus	O
Garry	O
Kasparov	O
match	O
.	O
</s>
<s>
Since	O
then	O
,	O
machine	O
learning	O
agents	B-General_Concept
have	O
shown	O
ever	O
greater	O
success	O
than	O
previous	O
AI	B-Application
agents	B-General_Concept
.	O
</s>
<s>
Go	O
is	O
another	O
turn-based	O
strategy	B-Application
game	I-Application
which	O
is	O
considered	O
an	O
even	O
more	O
difficult	O
AI	B-Application
problem	O
than	O
chess	B-Application
.	O
</s>
<s>
The	O
state	O
space	O
of	O
is	O
Go	O
is	O
around	O
10^170	O
possible	O
board	O
states	O
compared	O
to	O
the	O
10^120	O
board	O
states	O
for	O
Chess	B-Application
.	O
</s>
<s>
Prior	O
to	O
recent	O
deep	B-Algorithm
learning	I-Algorithm
models	O
,	O
AI	B-Application
Go	O
agents	B-General_Concept
were	O
only	O
able	O
to	O
play	O
at	O
the	O
level	O
of	O
a	O
human	O
amateur	O
.	O
</s>
<s>
Google	O
's	O
2015	O
AlphaGo	B-Application
was	O
the	O
first	O
AI	B-Application
agent	O
to	O
beat	O
a	O
professional	O
Go	O
player	O
.	O
</s>
<s>
AlphaGo	B-Application
used	O
a	O
deep	B-Algorithm
learning	I-Algorithm
model	O
to	O
train	O
the	O
weights	O
of	O
a	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
(	O
MCTS	O
)	O
.	O
</s>
<s>
The	O
deep	B-Algorithm
learning	I-Algorithm
model	O
consisted	O
of	O
2	O
ANN	O
,	O
a	O
policy	O
network	O
to	O
predict	O
the	O
probabilities	O
of	O
potential	O
moves	O
by	O
opponents	O
,	O
and	O
a	O
value	O
network	O
to	O
predict	O
the	O
win	O
chance	O
of	O
a	O
given	O
state	O
.	O
</s>
<s>
The	O
deep	B-Algorithm
learning	I-Algorithm
model	O
allows	O
the	O
agent	O
to	O
explore	O
potential	O
game	O
states	O
more	O
efficiently	O
than	O
a	O
vanilla	O
MCTS	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	I-Application
,	O
another	O
implementation	O
of	O
AlphaGo	B-Application
,	O
was	O
able	O
to	O
train	O
entirely	O
by	O
playing	O
against	O
itself	O
.	O
</s>
<s>
StarCraft	B-Application
and	O
its	O
sequel	O
StarCraft	B-Application
II	I-Application
are	O
real-time	O
strategy	O
(	O
RTS	O
)	O
video	O
games	O
that	O
have	O
become	O
popular	O
environments	O
for	O
AI	B-Application
research	I-Application
.	O
</s>
<s>
Blizzard	O
and	O
DeepMind	B-Application
have	O
worked	O
together	O
to	O
release	O
a	O
public	O
StarCraft	B-Application
2	I-Application
environment	O
for	O
AI	B-Application
research	I-Application
to	O
be	O
done	O
on	O
.	O
</s>
<s>
Various	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
have	O
been	O
tested	O
on	O
both	O
games	O
,	O
though	O
most	O
agents	B-General_Concept
usually	O
have	O
trouble	O
outperforming	O
the	O
default	O
AI	B-Application
with	O
cheats	O
enabled	O
or	O
skilled	O
players	O
of	O
the	O
game	O
.	O
</s>
<s>
Alphastar	B-Application
was	O
the	O
first	O
AI	B-Application
agent	O
to	O
beat	O
professional	O
StarCraft	B-Application
2	I-Application
players	O
without	O
any	O
in-game	O
advantages	O
.	O
</s>
<s>
The	O
deep	B-Algorithm
learning	I-Algorithm
network	O
of	O
the	O
agent	O
initially	O
received	O
input	O
from	O
a	O
simplified	O
zoomed	O
out	O
version	O
of	O
the	O
gamestate	O
,	O
but	O
was	O
later	O
updated	O
to	O
play	O
using	O
a	O
camera	O
like	O
other	O
human	O
players	O
.	O
</s>
<s>
The	O
developers	O
have	O
not	O
publicly	O
released	O
the	O
code	O
or	O
architecture	O
of	O
their	O
model	O
,	O
but	O
have	O
listed	O
several	O
state	O
of	O
the	O
art	O
machine	O
learning	O
techniques	O
such	O
as	O
relational	O
deep	O
reinforcement	O
learning	O
,	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
,	O
auto-regressive	O
policy	O
heads	O
,	O
pointer	O
networks	O
,	O
and	O
centralized	O
value	O
baseline	O
.	O
</s>
<s>
Alphastar	B-Application
was	O
initially	O
trained	O
with	O
supervised	O
learning	O
,	O
it	O
watched	O
replays	O
of	O
many	O
human	O
games	O
in	O
order	O
to	O
learn	O
basic	O
strategies	O
.	O
</s>
<s>
Dota	B-Application
2	I-Application
is	O
a	O
multiplayer	O
online	O
battle	O
arena	O
(	O
MOBA	O
)	O
game	O
.	O
</s>
<s>
Like	O
other	O
complex	O
games	O
,	O
traditional	O
AI	B-Application
agents	B-General_Concept
have	O
not	O
been	O
able	O
to	O
compete	O
on	O
the	O
same	O
level	O
as	O
professional	O
human	O
player	O
.	O
</s>
<s>
The	O
only	O
widely	O
published	O
information	O
on	O
AI	B-Application
agents	B-General_Concept
attempted	O
on	O
Dota	B-Application
2	I-Application
is	O
OpenAI	O
's	O
deep	B-Algorithm
learning	I-Algorithm
Five	O
agent	O
.	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
utilized	O
separate	O
LSTM	B-Algorithm
networks	O
to	O
learn	O
each	O
hero	O
.	O
</s>
<s>
It	O
trained	O
using	O
a	O
reinforcement	O
learning	O
technique	O
known	O
as	O
Proximal	O
Policy	O
Learning	O
running	O
on	O
a	O
system	O
containing	O
256	O
GPUs	B-Architecture
and	O
128,000	O
CPU	B-Architecture
cores	I-Architecture
.	O
</s>
<s>
It	O
was	O
eventually	O
able	O
to	O
beat	O
the	O
2018	O
Dota	B-Application
2	I-Application
esports	O
champion	O
team	O
in	O
a	O
2019	O
series	O
of	O
games	O
.	O
</s>
<s>
Planetary	B-Application
Annihilation	I-Application
is	O
a	O
real-time	O
strategy	B-Application
game	I-Application
which	O
focuses	O
on	O
massive	O
scale	O
war	O
.	O
</s>
<s>
The	O
developers	O
use	O
ANNs	O
in	O
their	O
default	O
AI	B-Application
agent	O
.	O
</s>
<s>
Supreme	B-Application
Commander	I-Application
2	I-Application
is	O
a	O
real-time	O
strategy	O
(	O
RTS	O
)	O
video	O
game	O
.	O
</s>
<s>
The	O
game	O
uses	O
Multilayer	B-Algorithm
Perceptrons	I-Algorithm
(	O
MLPs	O
)	O
to	O
control	O
a	O
platoon	O
’s	O
reaction	O
to	O
encountered	O
enemy	O
units	O
.	O
</s>
<s>
There	O
have	O
been	O
attempts	O
to	O
make	O
machine	O
learning	O
agents	B-General_Concept
that	O
are	O
able	O
to	O
play	O
more	O
than	O
one	O
game	O
.	O
</s>
<s>
These	O
"	O
general	O
"	O
gaming	O
agents	B-General_Concept
are	O
trained	O
to	O
understand	O
games	O
based	O
on	O
shared	O
properties	O
between	O
them	O
.	O
</s>
<s>
AlphaZero	B-Application
is	O
a	O
modified	O
version	O
of	O
AlphaGo	B-Application
Zero	I-Application
which	O
is	O
able	O
to	O
play	O
Shogi	O
,	O
chess	B-Application
,	O
and	O
Go	O
.	O
</s>
<s>
DeepMind	B-Application
was	O
able	O
to	O
train	O
this	O
generalized	O
agent	O
to	O
be	O
competitive	O
with	O
previous	O
versions	O
of	O
itself	O
on	O
Go	O
,	O
as	O
well	O
as	O
top	O
agents	B-General_Concept
in	O
the	O
other	O
two	O
games	O
.	O
</s>
<s>
Machine	O
learning	O
agents	B-General_Concept
are	O
often	O
not	O
covered	O
in	O
many	O
game	O
design	O
courses	O
.	O
</s>
<s>
Previous	O
use	O
of	O
machine	O
learning	O
agents	B-General_Concept
in	O
games	O
may	O
not	O
have	O
been	O
very	O
practical	O
,	O
as	O
even	O
the	O
2015	O
version	O
of	O
AlphaGo	B-Application
took	O
hundreds	O
of	O
CPUs	O
and	O
GPUs	B-Architecture
to	O
train	O
to	O
a	O
strong	O
level	O
.	O
</s>
<s>
This	O
potentially	O
limits	O
the	O
creation	O
of	O
highly	O
effective	O
deep	B-Algorithm
learning	I-Algorithm
agents	B-General_Concept
to	O
large	O
corporations	O
or	O
extremely	O
wealthy	O
individuals	O
.	O
</s>
<s>
The	O
extensive	O
training	O
time	O
of	O
neural	B-Architecture
network	I-Architecture
based	O
approaches	O
can	O
also	O
take	O
weeks	O
on	O
these	O
powerful	O
machines	O
.	O
</s>
<s>
AlphaStar	B-Application
shows	O
this	O
weakness	O
,	O
despite	O
being	O
able	O
to	O
beat	O
professional	O
players	O
,	O
it	O
is	O
only	O
able	O
to	O
do	O
so	O
on	O
a	O
single	O
map	O
when	O
playing	O
a	O
mirror	O
protoss	O
matchup	O
.	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
also	O
shows	O
this	O
weakness	O
,	O
it	O
was	O
only	O
able	O
to	O
beat	O
professional	O
player	O
when	O
facing	O
a	O
very	O
limited	O
hero	O
pool	O
out	O
of	O
the	O
entire	O
game	O
.	O
</s>
<s>
This	O
example	O
show	O
how	O
difficult	O
it	O
can	O
be	O
to	O
train	O
a	O
deep	B-Algorithm
learning	I-Algorithm
agent	O
to	O
perform	O
in	O
more	O
generalized	O
situations	O
.	O
</s>
<s>
Machine	O
learning	O
agents	B-General_Concept
have	O
shown	O
great	O
success	O
in	O
a	O
variety	O
of	O
different	O
games	O
.	O
</s>
<s>
However	O
,	O
agents	B-General_Concept
that	O
are	O
too	O
competent	O
also	O
risk	O
making	O
games	O
too	O
difficult	O
for	O
new	O
or	O
casual	O
players	O
.	O
</s>
<s>
These	O
highly	O
trained	O
agents	B-General_Concept
are	O
likely	O
only	O
desirable	O
against	O
very	O
skilled	O
human	O
players	O
who	O
have	O
many	O
of	O
hours	O
of	O
experience	O
in	O
a	O
given	O
game	O
.	O
</s>
<s>
Given	O
these	O
factors	O
,	O
highly	O
effective	O
deep	B-Algorithm
learning	I-Algorithm
agents	B-General_Concept
are	O
likely	O
only	O
a	O
desired	O
choice	O
in	O
games	O
that	O
have	O
a	O
large	O
competitive	O
scene	O
,	O
where	O
they	O
can	O
function	O
as	O
an	O
alternative	O
practice	O
option	O
to	O
a	O
skilled	O
human	O
player	O
.	O
</s>
<s>
Computer	B-Application
vision	I-Application
focuses	O
on	O
training	O
computers	O
to	O
gain	O
a	O
high-level	O
understanding	O
of	O
digital	O
images	O
or	O
videos	O
.	O
</s>
<s>
Many	O
computer	B-Application
vision	I-Application
techniques	O
also	O
incorporate	O
forms	O
of	O
machine	O
learning	O
,	O
and	O
have	O
been	O
applied	O
on	O
various	O
video	O
games	O
.	O
</s>
<s>
This	O
application	O
of	O
computer	B-Application
vision	I-Application
focuses	O
on	O
interpreting	O
game	O
events	O
using	O
visual	O
data	O
.	O
</s>
<s>
In	O
some	O
cases	O
,	O
artificial	B-General_Concept
intelligence	I-General_Concept
agents	I-General_Concept
have	O
used	O
model-free	B-Algorithm
techniques	O
to	O
learn	O
to	O
play	O
games	O
without	O
any	O
direct	O
connection	O
to	O
internal	O
game	O
logic	O
,	O
solely	O
using	O
video	O
data	O
as	O
input	O
.	O
</s>
<s>
Andrej	O
Karpathy	O
has	O
demonstrated	O
that	O
relatively	O
trivial	O
neural	B-Architecture
network	I-Architecture
with	O
just	O
one	O
hidden	O
layer	O
is	O
capable	O
of	O
being	O
trained	O
to	O
play	O
Pong	B-Application
based	O
on	O
screen	O
data	O
alone	O
.	O
</s>
<s>
In	O
2013	O
,	O
a	O
team	O
at	O
DeepMind	B-Application
demonstrated	O
the	O
use	O
of	O
deep	O
Q-learning	B-Algorithm
to	O
play	O
a	O
variety	O
of	O
Atari	O
video	O
games	O
Beamrider	B-Application
,	O
Breakout	B-Application
,	O
Enduro	B-Application
,	O
Pong	B-Application
,	O
Q*bert	B-Application
,	O
Seaquest	B-Application
,	O
and	O
Space	B-Application
Invaders	I-Application
from	O
screen	O
data	O
.	O
</s>
<s>
Doom	B-Application
(	O
1993	O
)	O
is	O
a	O
first-person	O
shooter	O
(	O
FPS	O
)	O
game	O
.	O
</s>
<s>
Student	O
researchers	O
from	O
Carnegie	O
Mellon	O
University	O
used	O
computer	B-Application
vision	I-Application
techniques	O
to	O
create	O
an	O
agent	O
that	O
could	O
play	O
the	O
game	O
using	O
only	O
image	O
pixel	O
input	O
from	O
the	O
game	O
.	O
</s>
<s>
The	O
students	O
used	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
CNN	B-Architecture
)	O
layers	O
to	O
interpret	O
incoming	O
image	O
data	O
and	O
output	O
valid	O
information	O
to	O
a	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
which	O
was	O
responsible	O
for	O
outputting	O
game	O
moves	O
.	O
</s>
<s>
Other	O
uses	O
of	O
vision-based	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
for	O
playing	O
games	O
have	O
included	O
playing	O
Super	B-Application
Mario	I-Application
Bros	I-Application
.	I-Application
only	O
using	O
image	O
input	O
,	O
using	O
deep	O
Q-learning	B-Algorithm
for	O
training	O
.	O
</s>
<s>
Researchers	O
with	O
OpenAI	O
created	O
about	O
2000	O
hours	O
of	O
video	O
plays	O
of	O
Minecraft	B-Application
coded	O
with	O
the	O
necessary	O
human	O
inputs	O
,	O
and	O
then	O
trained	O
a	O
machine	O
learning	O
model	O
to	O
comprehend	O
the	O
video	O
feedback	O
from	O
the	O
input	O
.	O
</s>
<s>
The	O
researchers	O
then	O
used	O
that	O
model	O
with	O
70,000	O
hours	O
of	O
Minecraft	B-Application
playthroughs	O
offered	O
on	O
YouTube	O
to	O
see	O
how	O
well	O
the	O
model	O
could	O
create	O
the	O
input	O
to	O
match	O
that	O
behavior	O
and	O
learn	O
further	O
from	O
it	O
,	O
such	O
as	O
being	O
able	O
to	O
learn	O
the	O
steps	O
and	O
process	O
of	O
creating	O
a	O
diamond	O
pickaxe	O
tool	O
.	O
</s>
<s>
PCG	O
has	O
been	O
used	O
in	O
various	O
games	O
for	O
different	O
types	O
of	O
content	O
generation	O
,	O
examples	O
of	O
which	O
include	O
weapons	O
in	O
Borderlands	B-Application
2	I-Application
,	O
all	O
world	O
layouts	O
in	O
Minecraft	B-Application
and	O
entire	O
universes	O
in	O
No	B-Application
Man	I-Application
's	I-Application
Sky	I-Application
.	O
</s>
<s>
Common	O
approaches	O
to	O
PCG	O
include	O
techniques	O
that	O
involve	O
grammars	O
,	O
search-based	B-Application
algorithms	I-Application
,	O
and	O
logic	B-Language
programming	I-Language
.	O
</s>
<s>
Machine	O
learning	O
techniques	O
used	O
for	O
content	O
generation	O
include	O
Long	B-Algorithm
Short-Term	I-Algorithm
Memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
Recurrent	B-Algorithm
Neural	I-Algorithm
Networks	I-Algorithm
(	O
RNN	O
)	O
,	O
Generative	B-Algorithm
Adversarial	I-Algorithm
networks	I-Algorithm
(	O
GAN	O
)	O
,	O
and	O
K-means	B-Algorithm
clustering	I-Algorithm
.	O
</s>
<s>
Not	O
all	O
of	O
these	O
techniques	O
make	O
use	O
of	O
ANNs	O
,	O
but	O
the	O
rapid	O
development	O
of	O
deep	B-Algorithm
learning	I-Algorithm
has	O
greatly	O
increased	O
the	O
potential	O
of	O
techniques	O
that	O
do	O
.	O
</s>
<s>
Galactic	B-Application
Arms	I-Application
Race	I-Application
is	O
a	O
space	O
shooter	O
video	O
game	O
that	O
uses	O
neuroevolution	B-Algorithm
powered	O
PCG	O
to	O
generate	O
unique	O
weapons	O
for	O
the	O
player	O
.	O
</s>
<s>
The	O
developers	O
use	O
a	O
form	O
of	O
neuroevolution	B-Algorithm
called	O
cgNEAT	O
to	O
generate	O
new	O
content	O
based	O
on	O
each	O
player	O
's	O
personal	O
preferences	O
.	O
</s>
<s>
Each	O
generated	O
item	O
is	O
represented	O
by	O
a	O
special	O
ANN	O
known	O
as	O
a	O
Compositional	B-Algorithm
Pattern	I-Algorithm
Producing	I-Algorithm
Network	I-Algorithm
(	O
CPPNs	B-Algorithm
)	O
.	O
</s>
<s>
During	O
the	O
evolutionary	O
phase	O
of	O
the	O
game	O
cgNEAT	O
calculates	O
the	O
fitness	O
of	O
current	O
items	O
based	O
on	O
player	O
usage	O
and	O
other	O
gameplay	O
metrics	O
,	O
this	O
fitness	O
score	O
is	O
then	O
used	O
decide	O
which	O
CPPNs	B-Algorithm
will	O
reproduce	O
to	O
create	O
a	O
new	O
item	O
.	O
</s>
<s>
Super	B-Application
Mario	I-Application
Bros	I-Application
.	I-Application
has	O
been	O
used	O
by	O
several	O
researchers	O
to	O
simulate	O
PCG	O
level	O
creation	O
.	O
</s>
<s>
PCG	O
level	O
creation	O
for	O
The	B-Application
Legend	I-Application
of	I-Application
Zelda	I-Application
has	O
been	O
attempted	O
by	O
researchers	O
at	O
the	O
University	O
of	O
California	O
,	O
Santa	O
Cruz	O
.	O
</s>
<s>
Machine	O
learning	O
has	O
seen	O
use	O
in	O
the	O
experimental	O
field	O
of	O
music	O
generation	O
;	O
it	O
is	O
uniquely	O
suited	O
to	O
processing	O
raw	O
unstructured	B-Application
data	I-Application
and	O
forming	O
high	O
level	O
representations	O
that	O
could	O
be	O
applied	O
to	O
the	O
diverse	O
field	O
of	O
music	O
.	O
</s>
<s>
Methods	O
include	O
the	O
use	O
of	O
basic	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
autoencoders	B-Algorithm
,	O
restricted	B-Algorithm
boltzmann	I-Algorithm
machines	I-Algorithm
,	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
generative	B-Algorithm
adversarial	I-Algorithm
networks	I-Algorithm
(	O
GANs	O
)	O
,	O
and	O
compound	O
architectures	O
that	O
use	O
multiple	O
methods	O
.	O
</s>
<s>
The	O
2014	O
research	O
paper	O
on	O
"	O
Variational	O
Recurrent	O
Auto-Encoders	B-Algorithm
"	O
attempted	O
to	O
generate	O
music	O
based	O
on	O
songs	O
from	O
8	O
different	O
video	O
games	O
.	O
</s>
<s>
The	O
neural	B-Architecture
network	I-Architecture
in	O
the	O
project	O
was	O
able	O
to	O
generate	O
data	O
that	O
was	O
very	O
similar	O
to	O
the	O
data	O
of	O
the	O
games	O
it	O
trained	O
off	O
of	O
.	O
</s>
