MSc project
Learning Behaviours of Autonomous Agents
Benoit Isaac
On this page you can find:
- My dissertation : Postcript
File, High Quality, 6.5 Mo (for double side printing)
or PDF, lower
quality, 1.8 Mo but with hyperlinks and bookmarks;
- The documentation
generated by Javadoc, with the source code;
- The program
itself, with instructions on how to test it - Quick download
(requires JRE > 1.4.2) : simuLCS.jar , 210 Ko.
Documentation
The documentation has been generated using Javadoc, and it is available
here.
It is linked with the source code, so for each class you can see the
source code by clicking on the appropriate link (see screenshot below).
Testing
the Program
The program is available as a Java archive (jar), so you need to
install the Java Run-Time Environment on your machine to
test it. It
was tested with the JRE 1.4.2, but more recent versions should work as
well. You can obtain the JRE from
Sun, for example.
The Java Archive can be downloaded here : simuLCS.jar
NB: The program may slightly differ from the screenshots in the thesis
because we made several small improvements to the GUI so that the
system can be
tried more easily by examiners.
The resolution of your screen should be at least 1024x768 to
obtain a normal display.
Once you
have downloaded the jar file, the program can be launched
ususally by double-clicking on it or by the command :
java
-jar simuLCS.jar
This should run the display, with by default the Experiment 3
described in the thesis. Run the experiment by clicking on "Start", type the
base name of the data file (say
myfile)
and watch the
learning system working by clicking on the Tab "Watch Learning" and
selecting the Agent 2. You should be able to see the real behaviour and
the expected
behaviour, with usually more and more "-EXACT-" rules over time. Click
on "Finish"
to stop the experiment.
Then you can run : gnuplot
myfile.gp to obtain a Postcript Graph of the Average
Reward. (This requires gnuplot to be installed on your machine, see www.gnuplot.info) .
The file myfile.rules recorded the population of rules inside the LCS at the end of the
experiment.
You need to click on "Reset"
to start another experiment. You can add/remove several Agents and
change their behaviours (by adding/removing classifiers) ; or add an AgentInteractive
and some AgentDucks
to check the flocking behaviour.
Finally, you can run experiment 3 without the display by adding an
argument to the command line, namely the base name of the output files,
ie. :
java
-jar simuLCS.jar myfile
This will launch successively 3 experiments (Experiment 3, one Agent 2,
one Agent Ra and the Arena) until 50 000 time steps. Run gnuplot
myfile*.gp to obtain all the Postscript graphs.
For any question or remark, please send me an e-mail to the address in the thesis, or use my contact form.