drozzy / reinforce_tianshou

Reinforce Tianshou

My implementation of REINFORCE RL algorithm with the help of tianshou reinforcement learning framework. This builds on top of my implementation using only PyTorch here: https://github.com/drozzy/reinforce

The code trains an agent to solve a CartPole-v0 environment and then renders a few episodes with a trained agent.

Video Explanation

See my video here: Tianshou: Best RL Framework You Never Heard Of

Contents:

1. The reinforce_tianshou.py is the complete example.
2. The intermediate_steps/reinforce_tianshou_no_trainer.py shows how things would look without a trainer.
3. The intermediate_steps/reinforce_tianshou_no_trainer_no_net.pyshows things without a trainer and a built-in network.
4. The intermediate_steps/reinforce_tianshou_no_net.py shows how to create a custom network and a custom policy, while using built-in trainer.
5. The slides_code/policy_component.py - shows an example of calling a built-in policy on an observation from CartPole environment.

Install

conda env create -f environment.yml
conda activate reinforce_tianshou
pip install -r requirements.txt

Run

python reinforce_tianshou.py

References

• p.328 of Reinforcement Learning 2nd Ed. Sutton & Barto.
• Reinforce with Pytorch https://github.com/drozzy/reinforce

--Andriy Drozdyuk