Truncated rollouts
Scott Steiner
could someone explain to me, in general, what truncated rollouts are?
Also, in gnubg there is a "Truncation" field in the "Rollouts" menu item
which is set at a default value of 7, hopefully I will understand how to
use this setting once I comprehend the theory behind truncated rollouts.
thx!
Jørn Thyssen
> Hi,
>
> could someone explain to me, in general, what truncated rollouts are?
In a non-truncated rollout all trials are played until the end. In a
truncated rollout you can specify that the rollout is truncated at, say,
7-ply, that is, each trial will only include at most 7 turns. This is of
course much faster than playing each trial till the end, but the price
may be lost precision or a wrong result of the rollout.
For example, a 1296 trial rollout truncated at 2 ply is equivalent to a
2-ply evaluation.
> Also, in gnubg there is a "Truncation" field in the "Rollouts" menu item
> which is set at a default value of 7, hopefully I will understand how to
> use this setting once I comprehend the theory behind truncated rollouts.
With the default setting the rollout is truncated at 7 ply. If you set
it to 0 gnubg will not truncate the rollout.
The optimum settings depend very much on the problem at hand: cube
decision or chequer play; holding game, prime-vs-prime, backgame etc etc.
Jørn
Gregg Cattanach
the bot's evaluation function at the position obtained to accumulate the
results that go into the rollout. For play vs. play comparisons
(especially where both sides have simple or difficult plays), truncated
rollouts probably give pretty reasonable results. *(Some caution is
required when one side has very simple plays and the other is very
difficult, like containment positions. In these the errors will pile up on
one side, and that can skew the results.) To get the same accuracy for full
rollouts that you get from a truncated rollout you usually have to do 40
times the number of games because a fully played out game has so much more
'noise' involved. And, of course, full rollouts take dramatically longer.
So given limited CPU resources, on checker play comparisons truncation is
probably quite useful and gives up only a small amount of accuracy. (I
usually truncate at 11 moves.)
For cube action rollouts, the absolute equity value is of critical
importance, so any error the bot has on estimating the absolute equity of
the resulting truncated position will find its way into the rollout results.
Most people would say that truncation for cube action is probably the wrong
way to go; full rollouts are required. Checker plays we're trying to answer
which play is best, and we're not too worried what the absolute value of the
best play is. For cube action the absolute equity is the only important
value.
Gregg C.
"Scott Steiner" <nos...@nospam.nospam> wrote in message
news:3DA94C2C...@nospam.nospam...