tomshlomo/cs234_hw2

RL with Atari

Install

Install either Anaconda or Miniconda using instructions below

https://docs.conda.io/projects/continuumio-conda/en/latest/user-guide/install/index.html

Environment

Pong-v0

• We play against a decent AI player.
• One player wins if the ball pass through the other player and gets reward +1 else -1.
• Episode is over when one of the player reaches 21 wins
• final score is between -21 or +21 (lost all or won all)

# action = int in [0, 6)
# state  = (210, 160, 3) array
# reward = 0 during the game, 1 if we win, -1 else

We use a modified env where the dimension of the input is reduced to

# state = (80, 80, 1)

with downsampling and greyscale.

Training

Once done with implementing q2_linear.py (setup of the tensorflow necessary op) and q3_nature make sure you test your implementation by launching python q2_linear.py and python q3_nature.py that will run your code on the Test environment.

You can launch the training of DeepMind's DQN on pong with

python q5_train_atari_nature.py

The default config file should be sufficient to reach good performance after 5 million steps.

Training tips: (1) The starter code writes summaries of a bunch of useful variables that can help you monitor the training process. You can monitor your training with Tensorboard by doing, on Azure

tensorboard --logdir=results

and then connect to ip-of-you-machine:6006

(2) You can use ‘screen’ to manage windows on VM and to retrieve running programs. Before training DQN on Atari games, run

screen 

then run

python q5_train_atari_nature.py

By using Screen, programs continue to run when their window is currently not visible and even when the whole screen session is detached from the users terminal.

To detach from your window, simply press the following sequence of buttons

ctrl-a d

This is done by pressing control-a first, releasing it, and press d

To retrieve your running program on VM, simply type

screen -r

which will recover the detached window.

Credits Assignment code written by Guillaume Genthial and Shuhui Qu. Assignment code updated by Haojun Li and Garrett Thomas