<s>
DeepStack	B-Application
is	O
an	O
artificial	B-Application
intelligence	I-Application
computer	O
program	O
designed	O
to	O
play	O
two-player	O
poker	O
,	O
specifically	O
heads	O
up	O
no-limit	O
Texas	B-Device
hold	I-Device
'	I-Device
em	I-Device
.	O
</s>
<s>
Instead	O
,	O
DeepStack	B-Application
uses	O
several	O
algorithmic	O
innovations	O
,	O
such	O
as	O
the	O
use	O
of	O
neural	B-Architecture
networks	I-Architecture
and	O
continual	O
resolving	O
.	O
</s>
<s>
At	O
the	O
core	O
of	O
the	O
program	O
is	O
the	O
use	O
of	O
neural	B-Architecture
networks	I-Architecture
for	O
determining	O
the	O
value	O
of	O
specific	O
card	O
combinations	O
.	O
</s>
<s>
The	O
program	O
uses	O
search	O
with	O
the	O
neural	B-Architecture
networks	I-Architecture
and	O
continual	O
resolving	O
to	O
ensure	O
strategy	O
found	O
at	O
each	O
step	O
is	O
consistent	O
with	O
the	O
strategy	O
used	O
in	O
previous	O
steps	O
.	O
</s>
<s>
The	O
search	O
procedure	O
uses	O
counterfactual	O
regret	O
minimization	O
to	O
iteratively	O
update	O
strategy	O
in	O
its	O
lookahead	O
tree	O
,	O
and	O
the	O
neural	B-Architecture
networks	I-Architecture
are	O
used	O
for	O
leaf	O
evaluation	O
.	O
</s>
<s>
In	O
a	O
study	O
completed	O
December	O
2016	O
,	O
DeepStack	B-Application
defeated	O
11	O
professional	O
poker	O
players	O
by	O
playing	O
44,000	O
hands	O
of	O
poker	O
.	O
</s>
<s>
Over	O
all	O
games	O
played	O
,	O
DeepStack	B-Application
won	O
49	O
big	O
blinds/100	O
(	O
always	O
folding	O
would	O
only	O
lose	O
75	O
bb/100	O
)	O
,	O
over	O
four	O
standard	O
deviations	O
from	O
zero	O
,	O
making	O
it	O
the	O
first	O
computer	O
program	O
to	O
beat	O
professional	O
poker	O
players	O
in	O
heads-up	O
no-limit	O
Texas	B-Device
hold'em	I-Device
poker	O
.	O
</s>
<s>
Concurrently	O
with	O
DeepStack	B-Application
,	O
a	O
competing	O
approach	O
from	O
Carnegie	O
Mellon	O
University	O
research	O
group	O
was	O
published	O
,	O
called	O
Libratus	O
.	O
</s>
<s>
Libratus	O
does	O
not	O
use	O
neural	B-Architecture
networks	I-Architecture
for	O
leaf	O
evaluation	O
.	O
</s>
<s>
Dara	O
O'Kearney	O
,	O
an	O
Irish	O
poker	O
professional	O
who	O
completed	O
456	O
hands	O
,	O
claimed	O
that	O
DeepStack	B-Application
played	O
in	O
a	O
style	O
similar	O
to	O
one	O
used	O
by	O
some	O
human	O
players	O
,	O
based	O
on	O
game	O
theory	O
.	O
</s>
