******************************************************* * Please accept our apologies if you receive multiple * * copies of this message due to cross-posting. * *******************************************************
Hi all, I am pleased to announce the release of our Keepaway benchmark player framework source code for use with the 2D soccer server used at RoboCup. In case you are unfamiliar with it, the keepaway domain is a subtask of soccer in which a team of "keepers" tries to maintain possession of the ball within a small playing region, while the opposing team of "takers" tries to take the ball away. This domain has been used successfully in the past as a machine learning testbed. See the following publications for more information: Peter Stone, Richard S. Sutton, and Gregory Kuhlmann. Scaling Reinforcement Learning toward RoboCup Soccer. http://www.cs.utexas.edu/~kuhlmann/papers/b2hd-AB2004-scaling.html Gregory Kuhlmann and Peter Stone. Progress in Learning 3 vs. 2 Keepaway. http://www.cs.utexas.edu/~kuhlmann/papers/b2hd-LNAI2003-keepaway.html Version 0.1 of the framework is available for download from our keepaway website: http://www.cs.utexas.edu/~AustinVilla/sim/keepaway/ The package contains the source code for the Keepaway benchmark player framework and the source code for some keepaway utility programs. The framework includes all low- and mid-level keepaway behaviors. A few example high-level policies are included. Not included is any learning code. However, the framework was designed to make it easy to insert your own learning code. Also included on our page are two tutorials. The first tutorial explains step-by-step how to install the players, run a simulation, and generate a graph of the learning curve. The second tutorial explains how to modify the players to include your own learning code. Please let us know if you have any questions. Also, we would like to hear if and how you are using this code. Cheers, Greg Kuhlmann and Peter Stone