In this project, we implement several reinforcement learning algorithms to try to get the best score at the OpenAI Gym's BipedalWalker environment.
Here are some instructions to run them:
First, execute the file script-requirements.sh to install the required libraries.
Then, report to the folder of the algorithm you want to execute:
Run dqn_discrete_walker.ipynb in a Google Colab environment with GPU enabled.
Weights for the network are provided in DQN_weights.dat, jsut change the path for loading / saving weights to skip training (currently the training saves results in Google Drive if you mount it to the notebook).
Run the Genetic.ipynb notebook with the genetic.py file being in the same folder.
For DDPG with standard replay buffer, run the DDPG_Standard_Buffer_Replay.ipynb notebook. Running it on GPU is recommended.
For DDPG with Prioritized Experience Replay, run the DDPG_Prioritized_Experience_Replay.ipynb notebook with the utils.py file being in the same folder.
The outcome of 2000-episode training can be seen on the video called DDPG_Video_Best_Parameters.mp4.
For PPO, run the PPO.ipynb notebook.
Here is an animation showing the performance of the agent after training with DDPG (after hyperparameter fine-tuning).
