From: Russell Power (rjpower_at_u.washington.edu)
Date: Fri Oct 17 2003 - 12:08:24 PDT
The paper describes the generation via genetic algorithms of a set of
competitive Robocode tanks and provides an evaluation of those techniques
against existing strategies.
I enjoyed how the research described in the paper took a somewhat novel
approach to the task of building a competitive tank - while I have seen
genetic search techniques applied in other contexts, in this case, the
philosophy behind the concept seemed particularly fitting - evolving by
battle the most competitive robot. The use and evaluation of a custom
description format (TableRex) for indicating tank behavior also shed some
light on the critical importance of a good description in order for genetic
searches to be effective.
While interesting, the paper suffers from the lack of well-planned
experimental verification. The actual data regarding the performance of his
generated bot's is dispersed throughout the performance section, making it
difficult to tell precisely how effective his technique was. There seemed
to be a bias in the evaluation towards the authors techniques, which, while
somewhat expected, makes it difficult for a reader to fully accept the
authors statements. Also, the author failed to give the reasoning behind
various important decisions - how he arrived at the idea of using TableRex,
for instance, seems like an interesting situation of its own.
Open topics of interest for the paper include investigation of the
performance effect of various alternate tank description formats, and the
comparison of the methods used in the paper to other search techniques (i.e.
how would beam-search/simulated annealing/... across the TableRex space
compare?).
This archive was generated by hypermail 2.1.6 : Fri Oct 17 2003 - 12:08:00 PDT