Skip to content

Latest commit

 

History

History
43 lines (34 loc) · 1.37 KB

README.md

File metadata and controls

43 lines (34 loc) · 1.37 KB
Raw

gym-sandbox

  • Customize your own gym environment to simulate any RL problem you want to solve.
  • Principle: don't introduce any complexity, focus on algo test!
  • Similar project: https://github.com/deepmind/pycolab

How to use

  1. git clone this repo
  2. cd into repo
  3. pip install -e .
  4. Switch to your algo, and import gym_sandbox ; gym.make("xxx-v0")
    • all xxx env are in env_list.py of root dir.
    • Please do note that now env requires jupyter notebook

Code structure

  • env_list.py >> contains all env you can use
  • envs >> core code of envs
  • test_algos >> demo algos to solve the env, note it's just demo!
  • test >> unittests

Philosophy

Divide and Conquer.

  • Divide complex real problem into clean focused sub-tasks, and conquer by each.

Pipeline.

  • Integrate your solutions of all sub-tasks.

Origin

We want to solve complex games, like StarCraft, SuperMario, etc. However, these games contains many sub-tasks, if we can't solve one, we can't solve the whole. So we must ensure we can solve each sub-task, and then we have the confidence to solve all.

Features

  • OpenAI Gym
    • 100% compatible
  • Bokeh
    • We use bokeh to render game state, which is clean, intuitive, and
  • Multi-agent
    • Inspired by MADDPG

Credit

OpenAI's universe-starter-agent is a very cool project, where there're many good point of engineering design.