<s>
OpenAI	B-Operating_System
Five	I-Operating_System
is	O
a	O
computer	B-Application
program	I-Application
by	O
OpenAI	O
that	O
plays	O
the	O
five-on-five	O
video	O
game	O
Dota	B-Application
2	I-Application
.	O
</s>
<s>
By	O
choosing	O
a	O
game	O
as	O
complex	O
as	O
Dota	B-Application
2	I-Application
to	O
study	O
machine	O
learning	O
,	O
OpenAI	O
thought	O
they	O
could	O
more	O
accurately	O
capture	O
the	O
unpredictability	O
and	O
continuity	O
seen	O
in	O
the	O
real	O
world	O
,	O
thus	O
constructing	O
more	O
general	O
problem-solving	O
systems	O
.	O
</s>
<s>
The	O
algorithms	O
and	O
code	O
used	O
by	O
OpenAI	B-Operating_System
Five	I-Operating_System
were	O
eventually	O
borrowed	O
by	O
another	O
neural	B-Architecture
network	I-Architecture
in	O
development	O
by	O
the	O
company	O
,	O
one	O
which	O
controlled	O
a	O
physical	O
robotic	O
hand	O
.	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
has	O
been	O
compared	O
to	O
other	O
similar	O
cases	O
of	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
playing	O
against	O
and	O
defeating	O
humans	O
,	O
such	O
as	O
AlphaStar	B-Application
in	O
the	O
video	O
game	O
StarCraft	B-Application
II	I-Application
,	O
AlphaGo	B-Application
in	O
the	O
board	O
game	O
Go	O
,	O
Deep	B-General_Concept
Blue	I-General_Concept
in	O
chess	B-Application
,	O
and	O
Watson	B-Application
on	O
the	O
television	O
game	O
show	O
Jeopardy	O
!	O
.	O
</s>
<s>
OpenAI	O
decided	O
to	O
use	O
Dota	B-Application
2	I-Application
,	O
a	O
competitive	O
five-on-five	O
video	O
game	O
,	O
as	O
a	O
base	O
due	O
to	O
it	O
being	O
popular	O
on	O
the	O
live	O
streaming	O
platform	O
Twitch	O
,	O
having	O
native	O
support	O
for	O
Linux	B-Application
,	O
and	O
had	O
an	O
application	B-Application
programming	I-Application
interface	I-Application
(	O
API	B-Application
)	O
available	O
.	O
</s>
<s>
After	O
the	O
match	O
,	O
CTO	O
Greg	O
Brockman	O
explained	O
that	O
the	O
bot	O
had	O
learned	O
by	O
playing	O
against	O
itself	O
for	O
two	O
weeks	O
of	O
real	B-General_Concept
time	I-General_Concept
,	O
and	O
that	O
the	O
learning	O
software	O
was	O
a	O
step	O
in	O
the	O
direction	O
of	O
creating	O
software	O
that	O
can	O
handle	O
complex	O
tasks	O
"	O
like	O
being	O
a	O
surgeon	O
"	O
.	O
</s>
<s>
At	O
The	O
International	O
2018	O
,	O
OpenAI	B-Operating_System
Five	I-Operating_System
played	O
in	O
two	O
games	O
against	O
professional	O
teams	O
,	O
one	O
against	O
the	O
Brazilian-based	O
paiN	O
Gaming	O
and	O
the	O
other	O
against	O
an	O
all-star	O
team	O
of	O
former	O
Chinese	O
players	O
.	O
</s>
<s>
Although	O
the	O
bots	O
lost	O
both	O
matches	O
,	O
OpenAI	O
still	O
considered	O
it	O
a	O
successful	O
venture	O
,	O
stating	O
that	O
playing	O
against	O
some	O
of	O
the	O
best	O
players	O
in	O
Dota	B-Application
2	I-Application
allowed	O
them	O
to	O
analyze	O
and	O
adjust	O
their	O
algorithms	O
for	O
future	O
games	O
.	O
</s>
<s>
Each	O
OpenAI	B-Operating_System
Five	I-Operating_System
bot	O
is	O
a	O
neural	B-Architecture
network	I-Architecture
containing	O
a	O
single	O
layer	O
with	O
a	O
4096-unit	O
LSTM	B-Algorithm
that	O
observes	O
the	O
current	O
game	O
state	O
extracted	O
from	O
the	O
Dota	O
developer	O
's	O
API	B-Application
.	O
</s>
<s>
The	O
neural	B-Architecture
network	I-Architecture
conducts	O
actions	O
via	O
numerous	O
possible	O
action	O
heads	O
(	O
no	O
human	O
data	O
involved	O
)	O
,	O
and	O
every	O
head	O
has	O
meaning	O
.	O
</s>
<s>
The	O
AI	B-Application
system	O
observes	O
the	O
world	O
as	O
a	O
list	O
of	O
20,000	O
numbers	O
and	O
takes	O
an	O
action	O
by	O
conducting	O
a	O
list	O
of	O
eight	O
enumeration	O
values	O
.	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
has	O
been	O
developed	O
as	O
a	O
general-purpose	O
reinforcement	O
learning	O
training	O
system	O
on	O
the	O
"	O
Rapid	O
"	O
infrastructure	O
.	O
</s>
<s>
By	O
2018	O
,	O
OpenAI	B-Operating_System
Five	I-Operating_System
had	O
played	O
around	O
180	O
years	O
worth	O
of	O
games	O
in	O
reinforcement	O
learning	O
running	O
on	O
256	O
GPUs	B-Architecture
and	O
128,000	O
CPU	B-Device
cores	O
,	O
using	O
Proximal	B-Algorithm
Policy	I-Algorithm
Optimization	I-Algorithm
,	O
a	O
policy	O
gradient	O
method	O
.	O
</s>
<s>
Prior	O
to	O
OpenAI	B-Operating_System
Five	I-Operating_System
,	O
other	O
AI	B-Application
versus	O
human	O
experiments	O
and	O
systems	O
have	O
been	O
successfully	O
used	O
before	O
,	O
such	O
as	O
Jeopardy	O
!	O
</s>
<s>
with	O
Watson	B-Application
,	O
chess	B-Application
with	O
Deep	B-General_Concept
Blue	I-General_Concept
,	O
and	O
Go	O
with	O
AlphaGo	B-Application
.	O
</s>
<s>
In	O
comparison	O
with	O
other	O
games	O
that	O
have	O
used	O
AI	B-Application
systems	O
to	O
play	O
against	O
human	O
players	O
,	O
Dota	B-Application
2	I-Application
differs	O
as	O
explained	O
below	O
:	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
observes	O
every	O
fourth	O
frame	O
,	O
generating	O
20,000	O
moves	O
.	O
</s>
<s>
By	O
comparison	O
,	O
chess	B-Application
usually	O
ends	O
before	O
40	O
moves	O
,	O
while	O
Go	O
ends	O
before	O
150	O
moves	O
.	O
</s>
<s>
Thus	O
,	O
playing	O
Dota	B-Application
2	I-Application
requires	O
making	O
inferences	O
based	O
on	O
this	O
incomplete	O
data	O
,	O
as	O
well	O
as	O
predicting	O
what	O
their	O
opponent	O
could	O
be	O
doing	O
at	O
the	O
same	O
time	O
.	O
</s>
<s>
By	O
comparison	O
,	O
Chess	B-Application
and	O
Go	O
are	O
"	O
full-information	O
games	O
"	O
,	O
as	O
they	O
do	O
not	O
hide	O
elements	O
from	O
the	O
opposing	O
player	O
.	O
</s>
<s>
Continuous	O
action	O
space	O
:	O
Each	O
playable	O
character	O
in	O
a	O
Dota	B-Application
2	I-Application
game	O
,	O
known	O
as	O
a	O
hero	O
,	O
can	O
take	O
dozens	O
of	O
actions	O
that	O
target	O
either	O
another	O
unit	O
or	O
a	O
position	O
.	O
</s>
<s>
The	O
OpenAI	B-Operating_System
Five	I-Operating_System
developers	O
allow	O
the	O
space	O
into	O
170,000	O
possible	O
actions	O
per	O
hero	O
.	O
</s>
<s>
By	O
comparison	O
,	O
the	O
average	O
number	O
of	O
actions	O
in	O
chess	B-Application
is	O
35	O
and	O
250	O
in	O
Go	O
.	O
</s>
<s>
Continuous	O
observation	O
space	O
:	O
Dota	B-Application
2	I-Application
is	O
played	O
on	O
a	O
large	O
map	O
with	O
ten	O
heroes	O
,	O
five	O
on	O
each	O
team	O
,	O
along	O
with	O
dozens	O
of	O
buildings	O
and	O
non-player	O
character	O
(	O
NPC	O
)	O
units	O
.	O
</s>
<s>
The	O
OpenAI	O
system	O
observes	O
the	O
state	O
of	O
a	O
game	O
through	O
developers’	O
bot	O
API	B-Application
,	O
as	O
20,000	O
numbers	O
that	O
constitute	O
all	O
information	O
a	O
human	O
is	O
allowed	O
to	O
get	O
access	O
to	O
.	O
</s>
<s>
A	O
chess	B-Application
board	O
is	O
represented	O
as	O
about	O
70	O
lists	O
,	O
whereas	O
a	O
Go	O
board	O
has	O
about	O
400	O
enumerations	O
.	O
</s>
<s>
OpenAI	B-Operating_System
Five	I-Operating_System
have	O
received	O
acknowledgement	O
from	O
the	O
AI	B-Application
,	O
tech	O
,	O
and	O
video	O
game	O
community	O
at	O
large	O
.	O
</s>
<s>
Chess	B-Application
player	O
Garry	O
Kasparov	O
,	O
who	O
lost	O
against	O
the	O
Deep	B-General_Concept
Blue	I-General_Concept
AI	B-Application
in	O
1997	O
,	O
stated	O
that	O
despite	O
their	O
losing	O
performance	O
at	O
The	O
International	O
2018	O
,	O
the	O
bots	O
would	O
eventually	O
"	O
get	O
there	O
,	O
and	O
sooner	O
than	O
expected	O
"	O
.	O
</s>
<s>
In	O
a	O
conversation	O
with	O
MIT	O
Technology	O
Review	O
,	O
AI	B-Application
experts	O
also	O
considered	O
OpenAI	B-Operating_System
Five	I-Operating_System
system	O
as	O
a	O
significant	O
achievement	O
,	O
as	O
they	O
noted	O
that	O
Dota	B-Application
2	I-Application
was	O
an	O
"	O
extremely	O
complicated	O
game	O
"	O
,	O
so	O
even	O
beating	O
non-professional	O
players	O
was	O
impressive	O
.	O
</s>
<s>
In	O
contrast	O
,	O
Motherboard	O
wrote	O
that	O
the	O
victory	O
was	O
"	O
basically	O
cheating	O
"	O
due	O
to	O
the	O
simplified	O
hero	O
pools	O
on	O
both	O
sides	O
,	O
as	O
well	O
as	O
the	O
fact	O
that	O
bots	O
were	O
given	O
direct	O
access	O
to	O
the	O
API	B-Application
,	O
as	O
opposed	O
to	O
using	O
computer	B-Application
vision	I-Application
to	O
interpret	O
pixels	O
on	O
the	O
screen	O
.	O
</s>
<s>
The	O
Verge	O
wrote	O
that	O
the	O
bots	O
were	O
evidence	O
that	O
the	O
company	O
's	O
approach	O
to	O
reinforcement	O
learning	O
and	O
its	O
general	O
philosophy	O
about	O
AI	B-Application
was	O
"	O
yielding	O
milestones	O
"	O
.	O
</s>
<s>
In	O
2019	O
,	O
DeepMind	B-Application
unveiled	O
a	O
similar	O
bot	O
for	O
Starcraft	B-Application
II	I-Application
,	O
AlphaStar	B-Application
.	O
</s>
<s>
Like	O
OpenAI	B-Operating_System
Five	I-Operating_System
,	O
AlphaStar	B-Application
used	O
reinforcement	O
learning	O
and	O
self-play	O
.	O
</s>
<s>
The	O
Verge	O
reported	O
that	O
"	O
the	O
goal	O
with	O
this	O
type	O
of	O
AI	B-Application
research	I-Application
is	O
not	O
just	O
to	O
crush	O
humans	O
in	O
various	O
games	O
just	O
to	O
prove	O
it	O
can	O
be	O
done	O
.	O
</s>
<s>
Instead	O
,	O
it	O
’s	O
to	O
prove	O
that	O
—	O
with	O
enough	O
time	O
,	O
effort	O
,	O
and	O
resources	O
—	O
sophisticated	O
AI	B-Application
software	I-Application
can	O
best	O
humans	O
at	O
virtually	O
any	O
competitive	O
cognitive	O
challenge	O
,	O
be	O
it	O
a	O
board	O
game	O
or	O
a	O
modern	O
video	O
game.	O
"	O
</s>
<s>
They	O
added	O
that	O
the	O
DeepMind	B-Application
and	O
OpenAI	O
victories	O
were	O
also	O
a	O
testament	O
to	O
the	O
power	O
of	O
certain	O
uses	O
of	O
reinforcement	O
learning	O
.	O
</s>
<s>
In	O
2018	O
,	O
they	O
were	O
able	O
to	O
reuse	O
the	O
same	O
reinforcement	O
learning	O
algorithms	O
and	O
training	O
code	O
from	O
OpenAI	B-Operating_System
Five	I-Operating_System
for	O
Dactyl	O
,	O
a	O
human-like	O
robot	O
hand	O
with	O
a	O
neural	B-Architecture
network	I-Architecture
built	O
to	O
manipulate	O
physical	O
objects	O
.	O
</s>
