DQN agent for Wimblepong

This project is a DQN agent able to play in the pong-v0 OpenAI Gym environment. The environment is used in the Reinforcement Learning course at Aalto University, Finland.

🔧 How to use

• Clone the repository: https://github.com/MattiaMolon/Atari-Pong-RL.
• Install the dependencies listed in requirements.txt.
• to train: check the train_basic.py to train the agent against simpleAI or train_hibrid.py to train the agent one game against simpleAI and one game against himself (This helps the agent to not overfit against simpleAI alone). Run both files with the flag --train True.
• to test: check the train_basic.py to test the agent against simoleAI or train_hibrid.py to test the agent against another agent. Run both the files without the flag --train True.

🤖 Agents

• The SimpleAI agent is an agent that uses the absolute ball and player positions to follow the ball and reflect it in random directions.
• The agent implemented in this repository consists in a DQN agent with experience replay and can be found in agent.py. The weights in ./weights/hibrid_tuned_best.ai can reach a winrate of 80% against simpleAI.

📜 Report

The report pdf_files/RL_final_project.pdf includes all the details about the implementation of the agent and the training procedures followed to reach the aformentioned scores.

🌍 Environment

Additional details on the environment used and how to interact with it are available in the official repository of the course.

🌠 Future works:

• implementation DQN
• implementation experience replay, buffers, and preprocessing
• implementation of PPO
• implementation of actor-critic