In the years 2000 - 2002 when I started vsoc the following results where documented.
||First trends for optimal values of mutation rate, steps per match and selection interval
Optimal feed forward net (A)
||The test uses multiple feed forward nets. They differ in the
number of layers and in the numbers of connections between
this layers. The number of connections is defined by the
probabillyity of connections between two arbitrary nodes of
Optimal Kick Out Factor (KOF)
||Watching previous simulations showed the effect, that
hitting the ball out of the field was a common task although it
is not advatntageous for soccer playing clients. This
behaviour was forced by the fact that hitting the ball
increases the fittnes of players. Specially if the ball lays
near the out line players start to increase their fittness by
hitting the ball out again and again.
Optimal mutation rate
||For these tests feed forward ANNs where used. They where all
initialized with random weights at the beginning of each run.
One population consists of 21 individuals. From these
individuals 6 are choosen to participate at a match. 3 in the
right and tree in the left team. They play a match of 200 steps.
After the match the following parameters are collected for
each player to generate a fitness valule.
Optimal Selection Interval
||The goal of this test was to determine the optimal selection
interval for test runs. This means how many matches should be
played before a new generation of players should be
generated. Of cours longer selection intervals result in
better performance because the diffrent players are oftener
tested. But long selection intervalls extend the
development time and should be avoided if possible. The
question is to find the shortest selection interval that
still leads to reasonable results.
Optimal Steps per Match
||The goal of this test was to determine the optimal steps per for
test runs. That means how many steps should should a match at
least have to determine reasonable results. A small number of
steps per match would increase the performance but the
results could lead in wrong directions where bigger numbers
of steps per match decrease the performance but you may trust
the results. The question is to find the smallest number of
steps per match that still leads to reasonable results.
- top -