This is the code for NUS ME5406 Project2. Our team members are: Zhiyao Ren, Jiawei Chen and Boxuan Men.
The goal of our project is to train the car to move forward on the straights, 45 degree bend, 90 degree bend and 135 degree bend using deep reinforcement learning algorithms. And the car can finish our randomly generated complete track.
Clone this repository and navigate to it in your terminal. Then run:
pip install -r requirements.txt
You need to change the turning_point in main.py before training or testing. The track should start in (400, 500) and the width and height of the track image should be at least 500 larger than the maximum values of x and y in turning_point. You can set --track_w and --track_h to change the width and height of the track image. You can choose to start your track horizontally or vertically.
You can use lib/create_track to help you visualize the track and try to manually move forward on it.
You can run the training or testing code by:
bash main.sh x
x means how many episodes you want in the training process. If you are running the testing process, you can choose any x.
You can set up this run in main.sh. You should set --type Train for training or set --type Test for testing. You can select the algorithm --algorithm DQN. We support DQN, Double DQN and Dueling DQN in our code. You can use --resume_checkpoint path/to/model.pt to load your model for testing or further training. You can also set some hyperparameters like --batch_size 32, --lr 5e-5 and --epsilon 0.95.
You can download our trained model [checkpoint]. You can randomly choose a turning_point in main.py. This is a model for DuelingDQN, please set following in main.sh:
--algorithm DuelingDQN --type Test --resume_checkpoint path/to/model.pt --track_h 1300 --track_w 2100