Hands-On Reinforcement Learning

Thomas DESGREYS

REINFORCE algorithm

Training

see reinforce_cartpole.py

The model is trained and as save as "reinforce_cartpole_best.pth" and the evolutions of loss and score (aka reward) through the episodes are shown below. These graphics point out the instability of this training algorithm. Although, with a bit of luck we end up with a model that reaches the max steps permitted by this gym environment (500 steps)

Evaluation

see evaluate_reinforce_cartpole.py

During evaluation, we get a 100% success rate for 100 trials.

Familiarization with a complete RL pipeline:

Application to training a robotic arm

Stable-Baselines3

see a2c_sb3_cartpole.py

Hugging Face Hub

Link to the trained model (cartpole)

Weights & Biases

Link to the wandb run (cartpole)

Full workflow with panda-gym

see a2c_sb3_panda_reach.py

As I couldn't make it work on my PC (difficulties to install panda-gym), I've used Google Colab.

see my notebook here (online) or directly a2c_sb3_panda_reach.ipynb

Link to the trained model (panda reach)

Link to the wandb run (panda reach)