thakorneyp11 / rl-dino-gym

Applying Reinforcement Learning to play "T-Rex Chrome Dino Game" with custom gym environment

Reinforcement Learning on DinoGame

Applying Reinforcement Learning to play "T-Rex Chrome Dino Game"

1. Features

1.1 Model Training

• training script: training_pipeline/run_training.py
• observation space: can select between image or coordinate inputs
• observation space: apply concatenation to allow the model to be able to look-back X steps before each predictions
• policy selection: CnnPolicy of image-inputs or MlpPolicy for coordinate-inputs
• save gameplay-videos while training
• training configs and metrics are stored in Weights & Biases

1.2 Model Evaluation

• inference script: training_pipeline/model_evaluation.py
• able to inference the model prediction results as a gameplay-video

2. Setups

2.1 Weights & Biases setup

3. Installation

3.1 Docker approach

3.2 Virtualenv approach

3.3 Main Dependencies

4. How to run

4.1 Run model training

4.2 Run model evaluation

rl-dino-gym's Issues

Option for choosing between image-type or coordinate-type obs_space

• Allow the user to select between images or coordinates as the observation space inputs to the model
• Config the type of model inputs from image_obs parameter when creating the environment
• Adjust the step() function to return the image information when image_obs is True

update UI to be more understandable (ref in comment)

source: https://github.com/wayou/t-rex-runner

Env wrapper for creating and saving gameplay-videos

• Create an environment wrapper to handle the process of creating and saving gameplay-videos while in the training process
• Save gameplay-videos only when that episode achieve new maximum reward score

Item images are called from local absolute paths

• Currently all item images are identified from the their local absolute paths
• Should find other way to reference each images without using local absolute paths