<s>
TD-Gammon	B-Application
is	O
a	O
computer	O
backgammon	B-Application
program	O
developed	O
in	O
1992	O
by	O
Gerald	O
Tesauro	O
at	O
IBM	O
's	O
Thomas	O
J	O
.	O
Watson	O
Research	O
Center	O
.	O
</s>
<s>
Its	O
name	O
comes	O
from	O
the	O
fact	O
that	O
it	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
net	I-Architecture
trained	O
by	O
a	O
form	O
of	O
temporal-difference	O
learning	O
,	O
specifically	O
TD-Lambda	O
.	O
</s>
<s>
TD-Gammon	B-Application
achieved	O
a	O
level	O
of	O
play	O
just	O
slightly	O
below	O
that	O
of	O
the	O
top	O
human	O
backgammon	B-Application
players	O
of	O
the	O
time	O
.	O
</s>
<s>
It	O
explored	O
strategies	O
that	O
humans	O
had	O
not	O
pursued	O
and	O
led	O
to	O
advances	O
in	O
the	O
theory	O
of	O
correct	O
backgammon	B-Application
play	O
.	O
</s>
<s>
During	O
play	O
,	O
TD-Gammon	B-Application
examines	O
on	O
each	O
turn	O
all	O
possible	O
legal	O
moves	O
and	O
all	O
their	O
possible	O
responses	O
(	O
two-ply	O
lookahead	B-Algorithm
)	O
,	O
feeds	O
each	O
resulting	O
board	O
position	O
into	O
its	O
evaluation	B-General_Concept
function	I-General_Concept
,	O
and	O
chooses	O
the	O
move	O
that	O
leads	O
to	O
the	O
board	O
position	O
that	O
got	O
the	O
highest	O
score	O
.	O
</s>
<s>
In	O
this	O
respect	O
,	O
TD-Gammon	B-Application
is	O
no	O
different	O
than	O
almost	O
any	O
other	O
computer	O
board-game	O
program	O
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
innovation	O
was	O
in	O
how	O
it	O
learned	O
its	O
evaluation	B-General_Concept
function	I-General_Concept
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
learning	O
algorithm	O
consists	O
of	O
updating	O
the	O
weights	O
in	O
its	O
neural	B-Architecture
net	I-Architecture
after	O
each	O
turn	O
to	O
reduce	O
the	O
difference	O
between	O
its	O
evaluation	O
of	O
previous	O
turns	O
 '	O
board	O
positions	O
and	O
its	O
evaluation	O
of	O
the	O
present	O
turn	O
's	O
board	O
position	O
—	O
hence	O
"	O
temporal-difference	O
learning	O
"	O
.	O
</s>
<s>
After	O
each	O
turn	O
,	O
the	O
learning	O
algorithm	O
updates	O
each	O
weight	O
in	O
the	O
neural	B-Architecture
net	I-Architecture
according	O
to	O
the	O
following	O
rule	O
:	O
</s>
<s>
|	O
is	O
a	O
"	O
learning	B-General_Concept
rate	I-General_Concept
"	O
parameter	O
.	O
</s>
<s>
Unlike	O
previous	O
neural-net	O
backgammon	B-Application
programs	O
such	O
as	O
Neurogammon	O
(	O
also	O
written	O
by	O
Tesauro	O
)	O
,	O
where	O
an	O
expert	O
trained	O
the	O
program	O
by	O
supplying	O
the	O
"	O
correct	O
"	O
evaluation	O
of	O
each	O
position	O
,	O
TD-Gammon	B-Application
was	O
at	O
first	O
programmed	O
"	O
knowledge-free	O
"	O
.	O
</s>
<s>
In	O
early	O
experimentation	O
,	O
using	O
only	O
a	O
raw	O
board	O
encoding	O
with	O
no	O
human-designed	O
features	O
,	O
TD-Gammon	B-Application
reached	O
a	O
level	O
of	O
play	O
comparable	O
to	O
Neurogammon	O
:	O
that	O
of	O
an	O
intermediate-level	O
human	O
backgammon	B-Application
player	O
.	O
</s>
<s>
Even	O
though	O
TD-Gammon	B-Application
discovered	O
insightful	O
features	O
on	O
its	O
own	O
,	O
Tesauro	O
wondered	O
if	O
its	O
play	O
could	O
be	O
improved	O
by	O
using	O
hand-designed	O
features	O
like	O
Neurogammon	O
's	O
.	O
</s>
<s>
Indeed	O
,	O
the	O
self-training	O
TD-Gammon	B-Application
with	O
expert-designed	O
features	O
soon	O
surpassed	O
all	O
previous	O
computer	O
backgammon	B-Application
programs	O
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
exclusive	O
training	O
through	O
self-play	O
(	O
rather	O
than	O
tutelage	O
)	O
enabled	O
it	O
to	O
explore	O
strategies	O
that	O
humans	O
previously	O
had	O
not	O
considered	O
or	O
had	O
ruled	O
out	O
erroneously	O
.	O
</s>
<s>
Its	O
success	O
with	O
unorthodox	O
strategies	O
had	O
a	O
significant	O
impact	O
on	O
the	O
backgammon	B-Application
community	O
.	O
</s>
<s>
TD-Gammon	B-Application
found	O
that	O
the	O
more	O
conservative	O
play	O
of	O
24-23	O
was	O
superior	O
.	O
</s>
<s>
Tournament	O
players	O
began	O
experimenting	O
with	O
TD-Gammon	B-Application
'	O
s	O
move	O
,	O
and	O
found	O
success	O
.	O
</s>
<s>
Backgammon	B-Application
expert	O
Kit	O
Woolsey	O
found	O
that	O
TD-Gammon	B-Application
'	O
s	O
positional	O
judgement	O
,	O
especially	O
its	O
weighing	O
of	O
risk	O
against	O
safety	O
,	O
was	O
superior	O
to	O
his	O
own	O
or	O
any	O
human	O
's	O
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
excellent	O
positional	O
play	O
was	O
undercut	O
by	O
occasional	O
poor	O
endgame	O
play	O
.	O
</s>
<s>
The	O
endgame	O
requires	O
a	O
more	O
analytical	O
approach	O
,	O
sometimes	O
with	O
extensive	O
lookahead	B-Algorithm
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
limitation	O
to	O
two-ply	O
lookahead	B-Algorithm
put	O
a	O
ceiling	O
on	O
what	O
it	O
could	O
achieve	O
in	O
this	O
part	O
of	O
the	O
game	O
.	O
</s>
<s>
TD-Gammon	B-Application
'	O
s	O
strengths	O
and	O
weaknesses	O
were	O
the	O
opposite	O
of	O
symbolic	B-General_Concept
artificial	I-General_Concept
intelligence	I-General_Concept
programs	O
and	O
most	O
computer	O
software	O
in	O
general	O
:	O
it	O
was	O
good	O
at	O
matters	O
that	O
require	O
an	O
intuitive	O
"	O
feel	O
"	O
but	O
bad	O
at	O
systematic	O
analysis	O
.	O
</s>
