The binaries in this directory are the 2005 Austin Villa Coach
Competition team from the University of Texas at Austin. It competed
and won first place in the RoboCup competition in Osaka, Japan, July
13 - 19, 2005. 

For full results visit:
http://staff.science.uva.nl/~jellekok/robocup/rc05/results/coach/index.html

The coach was created by Brad Knox, Gregory Kuhlmann, and Peter
Stone. Please contact Brad Knox at brad_knox@yahoo.com or Greg
Kuhlmann at kuhlmann@cs.utexas.edu.  Note that this software is
provided as-is and is unsupported.

There are two programs: PatternAnalyzer (for off-line anazlysis of
log files) and Coach (for on-line pattern recognition and
declaration). They were compiled under Debian linux 3.1.

The coach implementation was done almost completely from scratch, but
some code and ideas were taken from our previous 2004 entry.


**********************
Running the programs
**********************
Both PatternAnalyzer and Coach are started by the methods established
for the coach competition. PatternAnalyzer is executed by the shell
script offline-analyzer.sh, typing "offline-analyzer
/path/to/logs". The log directory must contain the .rcg logs and a
patterns.list file, which lists the names of the pattern-log files
with ".rcg" removed. Coach is executed by online-coach.sh, typing
"online-coach.sh host port numberOfPatterns".

PatternAnalyzer records data about the pattern and no-pattern .rcg
logs. It then analyzes the data and assigns weights to inidividual
components of the data. These weights are determined by a comparison
of the the pattern and no-pattern versions of each component. The
result of this analysis is serialized into a patternName.ser file.

After PatternAnalyzer finishes, offline-analyzer.sh calls the
supplemental program NormalizeM. NormalizeM runs a mock online phase
using the provided pattern logs (in patterns.list) to simulate
multiple online games. For each patternName.ser file, it computes
scores as Coach would (hjgher score means higher likelihood of it
being active). Each pattern's mean score is stored when it is
reserialized and later used by Coach to normalize its online score.

Coach deserializes the pattern.ser files and periodically assigns
scores to the patterns. A simplified explanation of how it determines
scores is: Coach gives more points if the data from the online game
more closely resembles data from the pattern log than data from the
no-pattern log. At cycle 800, the two highest ranked patterns are
declared present. At 2500 the remaining declarations are made from the
top patterns not already declared.

  
*******************
More information
*******************    

This coach binary and more information about it can be found on
the UT Austin Villa coach site:
http://www.cs.utexas.edu/users/AustinVilla/sim/2005/

We would appreciate an email if you start using these binaries for any
evaluations: brad_knox@yahoo.com, kuhlmann@cs.utexas.edu,
pstone@cs.utexas.edu.

