From: Daniel J. Klein (djklein_at_u.washington.edu)
Date: Sun Oct 19 2003 - 22:27:38 PDT
Summary:
This paper presents an overview of RoboCode and a the workings of a
successful Genetic Algorithmic approach to learning successful tanks.
Review:
Jacob Eisenstein starts his paper with a clear overview of RoboCode.
This allows readers who are unfamiliar with RoboCode to catch up and
provides a nice summary for those who have already seen RoboCode.
Next, the author presents his approach to designing a successful tank -
a genetic algorithm based on TableREX. TableREX is the author's own
deviant of REX, a scripting language. I am not familiar with REX, but
it sounds like TableREX was a required enhancement to make a genetic
algorithm possible.
With the basics of TableREX laid out, the author then describes his
encoding method. One thing I noted in this section is that it seems he
leads the GA quite a bit. He has looked at a number of hand coded tanks
and made similar functions and program flows available to his program.
Next, the author presents his the GA he used. I do not know terribly
much about GA, but it seems that he has picked some magic numbers out of
nowhere. I imaging a lot of testing went into determining numbers such
as the elitism rate of 2%, but this information is not available in this
short paper.
Overall, the author provides a nice summary of experimental results.
The results are well organized and clearly show what works and what does
not. Also, further insight is available in later sections.
I enjoyed this paper and am looking forward to designing my own robot!
This archive was generated by hypermail 2.1.6 : Sun Oct 19 2003 - 22:29:42 PDT