Upper Confidence Tree Planner for ATARI games
(1) compile the codes:
-- set up the path to your xitari emulator (ALE) and caffe-dev (CAFFE) in makefile
-- make
(2) run UCT on ATARI games
-- ./atariUCTPlanner -rom_path=${HOME}/subsystem/roms/ms_pacman.bin -depth=25 -num_traj=100 -save_data=true -save_path=output
-- depth and num_traj specify the planning depth and number of sampled trajectories
-- rom_path specify the game ROM
-- if save_data is true, then UCT planning data is saved in the folder specified by save_path. In the above example, the game frames are stored in output/frames folder, and actions are stored in output/act
(3) training using caffe-dev -- a sample trajectory data is stored in output.tar.gz. train_sample shows how to train a neural network given the sample data
-- step 1: generate pixel-wise mean: ./preprocess.sh (assume caffe-dev binaries are in your PATH)
-- step 2: train the neural network: ./train.sh