Implementation of Google's paper on playing atari games using deep learning in python.

Paper Authors: Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, Martin Riedmiller
Paper Link: http://arxiv.org/abs/1312.5602

This project presents an implementation of a model (based on the above linked paper) that successfully learns control policies directly from high-dimensional sensory input using reinforcement learning. The model is a convolutional neural network, trained with a variant of Q-learning, whose input is raw pixels and whose output is a value function estimating future rewards. This model is tested on variety of Atari and custom made games and its performance is compared with human players.

Dependencies:

• Python 2.7
• numpy
• Lasagne
• Theano
• matplotlib
• scipy
• Arcade Learning environment (for Atari games)
• pygame (for flappy bird and shooter)
• GPU with CC score of greater than or equal to 3 (refer http://deeplearning.net/software/theano/library/tensor/nnet/conv.html and https://developer.nvidia.com/cuda-gpus)

The model is trained for 2 Atari games - Space Invaders and Breakout. The model was trained for about 12-13 hrs and has achieved good performace that is consistent with the paper.

To run the agent:

python AtariGame-Breakout/tester.py --rom_file breakout.bin --play_games 10 --display_screen --load_weights breakout_models/dep-q-rmsprop-breakout99-epoch.pkl 

python AtariGame-SpaceInvaders/tester.py --rom_file space_invaders.bin --play_games 10 --display_screen --load_weights spaceinvaders_models/dep-q-rmsprop-space_invaders99-epoch.pkl

I have trained a plain vanilla Q learning (based on http://sarvagyavaish.github.io/FlappyBirdRL/) based agent where the agent gets information such as the x and y distance from the pipes to compare the performance of this game-specific model to a generalized model as described in the Google's paper. Training time is about 2-3 hrs.

To run the agent:

python FlappyQ/run_qvflappy.py

Similar to the Atari games, I have trained the same model with minor only minor modificaions to the parameters to play Flappy Bird - although the performance is not as good as the Q learning mode which had explicit game data - it still gets a decent average score of about 20-30.

To run the agent:

python FlappyBirdDQN/ftester.py --play_games 10 --display_screen --load_weights flappy_models/dep-q-flappy-60-epoch.pkl

This is a very simple game I made using pygame where the player controls a "spaceship" is tasked to dodge the incomming "meteoroids" and stay alive as long as possible. I also tried an (silly?) experiment where I trained different models wherein each model had agents with different degrees of control over the space ship and compared the performance of the same.

To run the agent with just 2 control setting (left and right):

python ShooterDQN/stester2.py --play_games 10 --display_screen --load_weights shooter_models/dep-q-shooter-nipscuda-8movectrl-99-epoch.pkl 

To run the agent with just 4 control setting (left, right, top and bottom):

python ShooterDQN/stester4.py --play_games 10 --display_screen --load_weights shooter_models/dep-q-shooter-nipscuda-4movectrl-99-epoch.pkl 

To run the agent with just 8 control setting (all directions):

python ShooterDQN/stester8.py --play_games 10 --display_screen --load_weights shooter_models/dep-q-shooter-nipscuda-2movectrl-80-epoch.pkl 

For all the below graphs, the X axis is the traning timeline and the Y axis the score funtion for each game.
(Note: scores in Shooter anf flappy bird have been modified (reward amplified) because the original +1 or -1 is not applicable since the player does not have "lives" here and rewards are also very sparse in the these 2 games.)

Atari Breakout:

Atari Space Invaders:

Flappy Q Learning:

Flappy DQN:

Shooter (4-control):

Atari Breakout:

Atari Space Invaders:

Flappy Bird - DQN:

Flappy Bird - Q Learning:

Shooter (custom game):

Number of epochs and train cycles has been adjusted such that all the above code when used for traning takes only about 12-15 hrs max. depending on your CPU and GPU (My CPU: i5 3.4 GHz and GPU: nVidia GeForce 660). Also, do not expect super human level performance (as said in Google's paper) from the models as I have trained it only for 12-15 hrs - more traning with further parameter tuning can improve the scores of all the above games.

The deep Q network used in this project is a modified version of spragunr's dqn code (https://github.com/spragunr/deep_q_rl).
[1] Deep Learning in Neural Networks: An Overview http://arxiv.org/abs/1404.7828
[2] The Arcade Learning Environment: https://www.jair.org/media/3912/live-3912-7087-jair.pdf
[3] ImageNet Classification with Deep Convolutional Neural Networks: https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
[4] Lasagne: https://github.com/Lasagne/Lasagne
[5] Theano: http://deeplearning.net/software/theano/
[6] CUDA: https://developer.nvidia.com/cudnn
[7] Pygame: http://www.pygame.org/docs/
[8] General: http://deeplearning.net/