From: Karthik Gopalratnam (karthikg_at_cs.washington.edu)
Date: Sun Oct 19 2003 - 23:40:16 PDT
"Evolving Robot Tank Controllers" - Jacob Eisenstein
This paper discusses the application of genetic programming to
design automatically evolving tank controllers for the RoboCode
environment.
Given that the author deals with designing robot controllers that
evolve based on genetic programming techniques, the main ideas in the
paper deal with formulating the problem in the genetic algorithms
framework: developing an efficient representation scheme, and an
effective evaluation strategy.
The paper presents a representation of robots in a REX-like language,
that achieves the twofold goal of representing different robot
controllers efficiently for the genetic algorithm, as well as enable
the evaluation of successive generations of controllers in an efficient
manner since the representation is merely interpreted by Java code, as
opposed to having to be compiled into bytecode for each evaluation.
Also presented are various methods of evaluating the effectiveness
of various controllers within the robocode framework.
The choice of fitness measure is not mathematically rigorous, ie.
there does not appear to be a formally quantifiable reason why the
evaluation measure used against a certain set of opponents should work
against an opponent that that controller has never encountered before.
The fact that the most effective learned robots did not learn to use
their guns to shoot is apparently an enourmous anomaly in this method.
Although the author makes a case for the fact that this seems to be an
anomaly because using the gun is merely intuitive from a human's
perspecitve, it seems that there are significant mathematically
quantifable gains to a controller that can use its weapon efectively at
least a small fraction of the time, since a single correct shot not
only diminishes the opponent's energy level, it also increases the
controller's energy.
This limitation seems to be a consequence of the fact that the
representation is circumscribed by the limitations of genetic
algorithms in general - there might not be an efficient way of
representing the trigonometric information required to evolve
controllers capable of predicting where the opponent is going to move,
and when and in what direction to shoot etc., while also being able to
maipulate these individual controllers in the context of mutations,
crossovers etc. Imroving the representation for genetic algorithms is
an important, open research question.
Another open research area that woudld be worth pursuing is to
determine whether learning better controllers is possible using neural
networks. The author makes a strong case for using neural networks to
overcome the difficulty of targeting, and the whether neural networks
combined with biiologically inspired learning techniques will generate
the best controllers remains to be seen.
This archive was generated by hypermail 2.1.6 : Sun Oct 19 2003 - 23:40:16 PDT