Sciborg

Old name: RLFramework

Reinforcement Learning Framework

Disclaimer

This project is still a work in progress. Everything here should work, but expect a high frequency of modifications and "improvements" over time.

TODOs (in this order)

• Add a basic RL minimal example on cartpole (similar to the one in goal-conditioned RL).
• Create a wiki, because this readme is becoming messy ... It should be precise enough to be used as a documentation.
• clean-up environments code (should not change their behaviour)
• Create new agents:
  • DIAYN: an agent that wrap an value-based RL agent and make it learn skills
  • A model-based agent (probably MuZero if I succeed ...)
  • ICM (intrinsic curiosity module)

Give it a try

run: PYTHONPATH=$PTHONPATH:. python3 experiments/goal_condtioned_rl/main.py

NB: run it from a terminal, it uses ANSI characters and will not be printed well if you run it from py-charm.

Installation

• Fork and clone the repository.
• or download it with wget:

PROJECT_NAME="your_new_rl_project" && \
    wget https://github.com/hbonnavaud/RLFramework/archive/master.zip && \
    unzip master.zip && \
    mv RLFramework-main $PROJECT_NAME && \
    rm master.zip

Using a venv

python3 -m venv venv
. venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt

Using the dockerfile

docker build . --tag rlf:latest
docker run -v "$(pwd):/rlf/" --gpus all rlf:latest python3 /rlf/experiments/goal_condtioned_rl/main.py

--gpus all allows you to run it using cuda. -v "$(pwd):/rlf/" mount the repository as "/rlf/" so it can be used by the docker container.

Usage

• Create your own project with a copy (download using wget) or a fork of this project.
• Add an experiment directory in /experiments and put your code inside.
• Import the agents and environments you need in your experiment directory (or other ones from gym ...).
• You can start with a copy of /experiments/goal_conditioned_rl/ which is an example (explained bellow).

/experiments/goal_conditioned_rl/ example

Parrallel simulations

The directory /experiments/goal_conditioned_rl/ contains a main.py file that launch simulations in parrallel from simulation.py, which launch a goal-conditioned simulation between an agent and an environment.

Extract and plot results

During the simulation, it stores its outputs in experiments/goal_conditioned_rl/outputs/<environment_name>/<agent_name>. Then plot_results.py read these outputs and generate a goal-reaching accuracy plot from them.

Save and reload

The simulation.py saves the agent and the ongoing simulation after each evaluation. So if your computer crash, you can re-launch the main.py, and the simulation.py will load the last simulation from outputs/ directory and continue it, even if you have a lot of simulations running in parrallel.

Discord notification

Once a simulation is done, if you put a discord webhook in simulation.py, it will send a message on discord channel associated with the webhook (you don't need to have a server for this, just create a discord server, and associate a channel with a webhook). The send_discord_message function can be imported from rlf/utils so it's not a big deal to use it in your own experiments.

Agents

< explanations incoming >

Environments

< explanations incoming >

Contribute

Feel free to open an issue on any subject and any question. The code in agents/, environments/ and utils/ aim to be implemented in order to be as much easy to use as possiple. If you have any idea about how to improve it, feel free to open an issue or to contact me, I will be glad to discuss about it and modify the code in consequence.

PRs for new agents and new environments are highly welcome (Fork + pull request only!).