This package provides a loosely-organized set of tools for feature generation using spectral graph theory on state spaces, primarily intended for value function approximation in reinforcement learning.

Craig Corcoran <ccor@cs.utexas.edu> Bryan Silverthorn <bcs@cargo-cult.org>

This package is implemented in Python. To begin, set up a virtual environment with virtualenv. From the project root (the directory containing this README):

python virtualenv.py --no-site-packages environment source environment/bin/activate

Then install the minimum set of relevant dependencies:

easy_install numpy easy_install scipy easy_install scikit-learn easy_install plac

The project experiments are organized as modules under the "specmine.experiments" package. To obtain, for example, the Tic-Tac-Toe eigenvector visualization data, run

python -m specmine.experiments.ttt_analyze_evs eigenvectors.csv

Please contact the authors if any questions or difficulties arise.