<s>
The	O
Pommerman	B-Application
Challenge	I-Application
is	O
a	O
multi-agent	O
game	O
to	O
test	O
autonomous	O
artificial	B-Application
intelligence	I-Application
systems	O
.	O
</s>
<s>
The	O
real-time	O
requirement	O
limits	O
the	O
use	O
of	O
compute-heavy	O
techniques	O
such	O
as	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
The	O
branching	B-Data_Structure
factor	I-Data_Structure
at	O
each	O
move	O
can	O
be	O
as	O
large	O
as	O
1	O
,	O
296	O
,	O
because	O
all	O
four	O
agents	O
act	O
in	O
each	O
step	O
,	O
choosing	O
among	O
six	O
possibilities	O
.	O
</s>
<s>
Explosions	O
derail	O
tree	B-Algorithm
search	I-Algorithm
techniques	O
,	O
as	O
searchews	O
with	O
less	O
than	O
10	O
levels	O
ignore	O
explosions	O
while	O
deeper	O
searches	O
consider	O
too	O
many	O
choices	O
(	O
given	O
the	O
branching	B-Data_Structure
factor	I-Data_Structure
)	O
.	O
</s>
<s>
A	O
hybrid	O
approach	O
uses	O
a	O
limited-depth	O
tree	B-Algorithm
search	I-Algorithm
followed	O
by	O
exploring	O
a	O
deterministic/pessimistic	O
scenario	O
.	O
</s>
<s>
The	O
3	O
best	O
performing	O
solutions	O
used	O
online	O
tree	B-Algorithm
search	I-Algorithm
.	O
</s>
