From: Xu Miao (xm_at_u.washington.edu)
Date: Sun Oct 19 2003 - 21:55:18 PDT
Title: Evolving Robot Rank Controllers
Author: Jacob Eisenstein
Reviewer: Xu Miao
Summary:
This paper described a successful application of genetic programming
to evolve a robocontroller.
Some important ideas:
1. Proper representation. The most important reason for the
success, as claimed by author himself, is the proper representation. Instead
of using Java bytecode, he used an advanced language TableREX, which is more
suitable to GP and can be interpreted by linear time. Moreover, it
simplified the search space(if we consider the GA as a beam-search problem)
so that a small population like 500 can achieve the evolution. Of course,
his fitness scaling technique also played an important role.
2. Subsumption paradigm is another important idea of his work.
With this paradigm, a java base program can real-time call the evolved
program based on the priority of event, which provided a very fast interface
between TableREX and Java. At the same time, the author made a little bit
different change on subsumption paradigm to allow the module communication
and collaboration.
Some flaws:
1. Dodging beats Targeting. Most of the evolved bots favor
dodging bullets instead of targeting and shooting. One of the reasons could
be TableREX lacks trigonometric functions, which made evolution ignored the
methods to correctly targeting. Another could be the fitness function didn't
pay too much attention to the reward of the high score. At the same time,
some profound reasons haven't been found. Maybe the nature of GA or maybe
some laws behind the intelligence we haven't known.
2. Overfitting beats Generalization. The result shows a
sensitivity of starting position, which means the strategies evolved are
more starting position orientated. The generalization is not good enough.
But with common sense, a really good and intelligent robot should be able to
beat the adversary at any starting position.
Possible research questions:
Same as the flaws
1. Develop a robot with targeting system. I found that most of
the strongest hand-coded robots are advanced targeting robots, they all
beats the squigbot and much more advanced than "showcase" and "starter" set.
Obviously an active targeting behavior is more close to the real
intelligence so that the research on this question could touch or even
answer some basic questions in AI.
2. Overcome overfitting. Either do more samples on
One-Adversary-Multi-Startingpoints cases, or do generalization by several
bots evolved on One-Adversary-One-StartingPoints cases. Anyway, a robot can
deal with novel situation means more intelligent.
This archive was generated by hypermail 2.1.6 : Sun Oct 19 2003 - 21:55:01 PDT