Original Paper: https://arxiv.org/pdf/1312.5602.pdf
Our Review: https://drive.google.com/file/d/1H-jJGkWA-M2kUMFBSNuFxQU3BJ6rujxg/view?usp=sharing
To stay true to the DeepMind paper, we implemented their Deep Q-learning method with the same convolutional neural network (CNN) architecture for state-value function approximation. Furthermore, the actual agent itself is a Deep Q-learning agent (DQA) that receives the 84x84 images of the Atari game as inputs and utilizes the network to make decisions on what actions to perform.
To run all dependencies run the code below
pip install -r requirements.txtIf any issues occur, make sure all these dependencies are installed
- pip install gym[atari]
- pip install gym[accept-rom-license]
- pip install numpy
- pip install tensorflow
- pip install matplotlib
- pip install opencv-python
- pip install pickle
- pip install tqdm
- pip install seaborn
- pip install scipy
- pip install pandasTo train our model, run the train_brick_breaker.py. You can configure the agent with different parameters. The list below are the parameters we used during training
| Parameters | Configuration |
|---|---|
| game | - |
| model | - |
| gamma | 0.99 |
| epsilon | 1 |
| epsilon_decay | 0.9/500_000 |
| replay_memory_size | 500_000 |
| exploration_steps | 100_000 |
| target_update_horizon | 10_000 |
| main_model_train_horizon | 4 |
| min_replay_memory_size | 32 |
| save_frequency | 250 |
We have 2 different models that can be evaluated ours and the transfered weights one. They can be run with these calls:
python3 evaluate_brick_breaker.py --model {ours,transfer} [--games GAMES] [--render]For example, to evaluate our model and get the average reward for 5 games while rendering each game:
python3 evaluate_brick_breaker.py --model ours --games 5 --render