Research Playground built on top of OpenAI's Atari Gym, prepared for implementing various Reinforcement Learning algorithms.

It can emulate any of the following games:

['Asterix', 'Asteroids', 'MsPacman', 'Kaboom', 'BankHeist', 'Kangaroo', 'Skiing', 'FishingDerby', 'Krull', 'Berzerk', 'Tutankham', 'Zaxxon', 'Venture', 'Riverraid', 'Centipede', 'Adventure', 'BeamRider', 'CrazyClimber', 'TimePilot', 'Carnival', 'Tennis', 'Seaquest', 'Bowling', 'SpaceInvaders', 'Freeway', 'YarsRevenge', 'RoadRunner', 'JourneyEscape', 'WizardOfWor', 'Gopher', 'Breakout', 'StarGunner', 'Atlantis', 'DoubleDunk', 'Hero', 'BattleZone', 'Solaris', 'UpNDown', 'Frostbite', 'KungFuMaster', 'Pooyan', 'Pitfall', 'MontezumaRevenge', 'PrivateEye', 'AirRaid', 'Amidar', 'Robotank', 'DemonAttack', 'Defender', 'NameThisGame', 'Phoenix', 'Gravitar', 'ElevatorAction', 'Pong', 'VideoPinball', 'IceHockey', 'Boxing', 'Assault', 'Alien', 'Qbert', 'Enduro', 'ChopperCommand', 'Jamesbond']

Check out corresponding Medium article: Atari - Reinforcement Learning in depth (Part 1: DDQN)

The ultimate goal of this project is to implement and compare various RL approaches with atari games as a common denominator.

1. Clone the repo.
2. Go to the project's root folder.
3. Install required packagespip install -r requirements.txt.
4. Launch atari. I recommend starting with help command to see all available modes python atari.py --help.

Deep Convolutional Neural Network by DeepMind

After 5M of steps (~40h on Tesla K80 GPU or ~90h on 2.9 GHz Intel i7 Quad-Core CPU):

Training:

Normalized score - each reward clipped to (-1, 1)

Testing:

Human average: ~372

DDQN average: ~479 (128%)

Training:

Normalized score - each reward clipped to (-1, 1)

Testing:

Human average: ~28

DDQN average: ~62 (221%)

Training:

Normalized score - each reward clipped to (-1, 1)

Testing:

Human average: ~29,000

GE average: 31,000 (106%)

Greg (Grzegorz) Surma

PORTFOLIO

GITHUB

BLOG