Basics
OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms.
For reinforcement learning problem, two part:
- the environment
- the agent
The core gym interface is Env, which is the unified environment interface. There is no interface for agents:
- reset(self): Reset the environment’s state. Returns observation.
- step(self, action): Step the environment by one timestep. Returns observation, reward, done, info.
- render(self, mode=‘human’): Render one frame of the environment. The default mode will do something human friendly, such as pop up a window.
Supported Environment
Included Environment
- Algorithmic
- Atari
- Box2d
- Classic control
- Robotics
- MuJoCo
- Toy Text
OpenAI Environment
- Procgen
- Gym-Retro
Third Party Environments
- PyBullet Robotics Environments
- Obstacle Tower
- gym-maze
- gym-gazebo
- osim-rl
- gym-minigrid & gym-miniworld
- gym-sokoban
- gym-duckietown
- gymfc
- gym-anytrading
- GymGo
- gym-electric-motor
Creating Your Own Environment
The most important thing is to keep the architecture of your package like this:
gym-foo/
README.md
setup.py
gym_foo/
__init__.py
envs/
__init__.py
foo_env.py
foo_extra_hard_env.py
See more in Create new environment for Gym.
Wrappers
Wrappers are list of general purpose wrappers for environments. These can perform pre/postprocessing on the data that is exchanged between the agent and the environment.
There are many wrappers which you can use to transform the observation and/or action spaces and do some other operations.
Agents
An “agent” describes the method of running an RL algorithm against an environment in the gym.
- RandomAgent： This simple agent leverages the environments ability to produce a random valid action and does so for each step.
- Cross-Enrorpy Agent
- DQN
- AgentNet
- rllab
- keras-rl
Note: Cover Picture