Playing Chrome's Dinosaur Game with DQN

Introduction

This project implements and trains an agent to play the Dinosaur Game using Deep Q Network (DQN). The network architecture for DQN to train the agent is similar to what proposed in the original paper.

The project also implements a custom gymnasium (previously OpenAI gym) environment for training a reinforcement learning agent to play the Dinosaur Game. For more details, see the envs.py file.

Note: this environment is a simplified replication of the Dinosaur Game, hence missing some features/game mechanisms.

Setup

The project is implemented in Python and it uses Poetry for package management.

git clone https://github.com/aome510/chrome-dino-game-rl.git
cd chrome-dino-game-rl
poetry install

Then run poetry shell to spawn a shell within the virtual environment containing the project's dependencies.

Note: this project specifies Python version 3.11.*. To use another Python version, please update pyproject.toml accordingly.

Usage

• python train.py to train the DQN model. Changes the parameters of the Trainer class to train the agent differently.
• python play.py human to play the game manually with keyboard.
• python play.py ai -m model.pth to play the game with an agent (using the parameters from model.pth).
• python experiment.py to run experiments with the trained agent. Refer to the file for more details on its usage.

Results

The agent is trained in 1000 episodes. The plot displaying the agent's results after x00 episodes:

The reward is defined to be the number of obstacles that the agent passes before colliding one.

The model's best parameters (based on the result plot) can be found in the model.pth file.

Examples

Example plays of the agent after 100,300,500,700,900 episodes in the same environment:

• after 100 episodes
• after 300 episodes
• after 500 episodes
• after 700 episodes
• after 900 episodes

Acknowledgement

The images in the assets folder are from the Chrome-Dino-Runner repo.