AI research environment for the Atari 2600 games 🤖.

gsurma gsurma Last update: Nov 01, 2023

Atari

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)

Purpose

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

Usage

  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.

DDQN

Hyperparameters

* GAMMA = 0.99
* MEMORY_SIZE = 900000
* BATCH_SIZE = 32
* TRAINING_FREQUENCY = 4
* TARGET_NETWORK_UPDATE_FREQUENCY = 40000
* MODEL_PERSISTENCE_UPDATE_FREQUENCY = 10000
* REPLAY_START_SIZE = 50000
* EXPLORATION_MAX = 1.0
* EXPLORATION_MIN = 0.1
* EXPLORATION_TEST = 0.02
* EXPLORATION_STEPS = 850000

Model Architecture

Deep Convolutional Neural Network by DeepMind

* Conv2D (None, 32, 20, 20)
* Conv2D (None, 64, 9, 9)
* Conv2D (None, 64, 7, 7)
* Flatten (None, 3136)
* Dense (None, 512)
* Dense (None, 4)

Trainable params: 1,686,180

Performance

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

SpaceInvaders

Training:

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

Testing:

Human average: ~372

DDQN average: ~479 (128%)


Breakout

Training:

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

Testing:

Human average: ~28

DDQN average: ~62 (221%)

Genetic Evolution

Atlantis

Training:

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

Testing:

Human average: ~29,000

GE average: 31,000 (106%)

Author

Greg (Grzegorz) Surma

PORTFOLIO

GITHUB

BLOG

Subscribe to our newsletter