ESM 5984 Project Contract for Doug G. Gaff

Principal Investigator:

Doug G. Gaff

phone: 231-5613



A Presentation of Simulation Results for Autonomous Robot Behavior Using a Learning Classifier System Control Paradigm.


The Learning Classifier System (LCS) is a rule-based, message-passing, artificial intelligence paradigm which provides a stimulus-response control algorithm. The LCS is unique in that it provides real-time control but also has the ability to learn new behaviors. However, even though the concept is over two decades old, very little research has been done in the area of applying learning classifier systems to autonomous robots. Part of my thesis research involves the application of a single LCS to a single robot in an environment scenario known as the "animat problem." In the animat problem, an agent (robot) attempts to learn behaviors that allow it to move around obstacles in an infinite plane while attempting to reach a goal. However, the typical LCS has a minimum of twelve parameters that can be adjusted to optimize its performance in a given application (such as the animat problem). Three of these parameters directly affect the rate at which the LCS learns, and it would be nice to offer some sort of characterization of the effect of varying these three parameters.

Progress (as of 20 February 1995):

I have written a C++ program which provides an LCS object and an environment in which to test the animat problem. This simulation software has been completed. In addition, I have run several simulations to determine general performance of the LCS. All of my data sets for this project will come from this simulation software. (I estimate that a typical 4-space data set will require about four days of continuous simulation to complete).


My primary goal is to examine various combinations of the three environment parameters with the hope that I can locate a local minimum, one that has thus far eluded me. This minimum is "local" because I will have to hold the other 9+ parameters constant while varying these three parameters. If no single local minimum exists (a very likely situation), I hope to find a trend in the data. Whatever my findings, I will summarize the results into a brief multimedia presentation that I will use in my thesis defense. As an indirect benefit of this project, I also hope to establish a visualization method that can be used by future researches to examine the complicated LCS parameter optimization problem.


I intend to use PV-Wave and Spyglass Dicer to view and analyze the resulting four-space data set. Since all of my data will originate from my C++ code, I do not anticipate a need to use AVS (unfortunately). I will prepare my multimedia demonstration using Macromedia Director.