This is a repo for demonstrating policy proximal optimization (PPO) on a few openai gym reinforcement learning (RL) environments at The Australian Institute for Machine Learning (AIML) https://www.adelaide.edu.au/aiml/

The code can run in its own environment but so that it is reproducible it is recommended to use the docker image.

There are two ways to use this repo, 1. create a venv and run it or 2. use the docker container (recommended)

Install swig: sudo apt-get install swig

• Create and environment: conda env create -f environment.yml
• Activate it: conda activate rl_demos

• Install Docker https://docs.docker.com/get-docker/
• Build the image rl_demos/demos/docker-compose build

To run: python rl_demos/demos/rl_demo --gym_env {environment name} --demo_type {discrete/continuous}

i.e. to run the Acrobot: python rl_demos/demos/rl_demo --gym_env Acrobot-v1 --demo_type discrete

To change demos, edit the relevant config file and change the type i.e. Acrobot-v1

(Docker) To run:

1. First clone this repo
2. Build the dockerfile
3. run docker-compose up

Different environments can be run by changing the command args in the docker-compose.yml

Note: To kill the demo hit 'q'

Working demos:

• Acrobot-v1: https://www.gymlibrary.dev/environments/classic_control/acrobot/
• CartPole-v1: https://www.gymlibrary.dev/environments/classic_control/cart_pole/
• MountainCar-v0: https://www.gymlibrary.dev/environments/classic_control/mountain_car/

Working demos:

• BipedalWalker-v3: https://www.gymlibrary.dev/environments/box2d/bipedal_walker/
• LunarLanderContinuous-v2: https://www.gymlibrary.dev/environments/box2d/lunar_lander/
• MountainCarContinuous-v0: https://www.gymlibrary.dev/environments/classic_control/mountain_car_continuous/