From: Masaharu Kobashi (mkbsh_at_cs.washington.edu)
Date: Sun Oct 19 2003 - 18:15:30 PDT
Paper title: Evolving Robot Tank Controllers by Jacob Eisenstein
This paper discusses the technique and its results of applying
the genetic algorithm to creating an adaptive robotic tank.
The most important ideas:
First, it incorporated the genetic algorithm in the learning process
which is rare according to the author among the existing Robocode robotic
tanks.
Second, he invented a nice architecture, TableRex, to convert the learning
result into the robot's program. It is one of the most difficult tasks in
implementing the learned experience into the parts of the existing
program.
The largest flaws in the paper:
The paper does not delve into the cause and effect analysis of
the relations between the learning environment, the the learning
mechanism and the resulting performance changes.
I am most interested in how the author's robot came to be
par with reportedly the strongest hand-coded robot in terms of
the detail factors which have been adapted through the learning.
But the question is not satisfactorily answered.
Two important open research questions:
One important factor which critically affects the robot's performance
is targeting. As author admits, it is very difficult to implement
effectively into the robot and there still exist a lot of possibilities
to apply new techniques in this respect.
Another important open question is a use of other techniques such as
neural net. Although the author cites a single instance of failure
of a neural net implementation in the robocode, it seems potentials of
other learning methods in the Robocode environment is still an open
question which is worth serious investigation.
This archive was generated by hypermail 2.1.6 : Sun Oct 19 2003 - 18:15:31 PDT