Note: this code starts directly from Dr.Nathan Sprague's implementation of DQN. His codebase at https://github.com/spragunr/deep_q_rl.

Required packages

1. Cython -http://cython.org/#download

2. Theano -if later you get an error "cannot import host_from_gpu" you may need to update theano, see http://deeplearning.net/software/theano/install_ubuntu.html.

3. Pylearn2 -http://deeplearning.net/software/pylearn2/

4. ALE (with replaced controller file) -Download ALE at http://www.arcadelearningenvironment.org/downloads/ -unzip -copy file in ale_files to replace ale_0.4.4/ale_0_4/src/controllers/rlglue_controller.cpp -cd ale_0.4.4/ale_0_4 -cp makefile.unix makefile (use makefil.mac if you have mac) -make

5. rl-glue core -https://code.google.com/p/rl-glue-ext/wiki/RLGlueCore

6. rl-glue python codec -https://code.google.com/p/rl-glue-ext/wiki/Python

Required changes:

1. cd src && vi ale_run.py change ROM, ALE, GLUE paths to be your own locations GLUE path should be like /u/[usrname]/lib/bin/rl_glue ALE path should be like /[...]/ale_0.4.4/ale_0_4/ale ROM path is wherever you put your rom files

To run the experiment: python ale_run.py

Plot results: python plot_results.py [data folder path]/results.csv

Watch performance: python ale_run_watch.py [data folder path]/network_file_99.pkl