shawx825 / luxai

• To run imitation learning, use nohup python main.py under working directory imitaion_learning. Results will be saved under fold model_checkpoint along with model checkpoints. The fold will be named after the current time-stamp. You can also tweak command arguments to customize training phase. For example,

  • nohup python main.py --load_prop 1.0 --epoch 30 --option 3 --map 12 --device 'cuda:0' > IL_OPT3_MAP12_TB.log 2>&1 &

  For details of or to modify command arguments, please see imitation_learning/main.py

• Code structures, configuration, command arguments, etc remain similar for Deep Q-learning.

• To submit via kaggle API, you need to ensure your API token file exists in ~/.kaggle/kaggle.json. Then run command:

  • kaggle competitions submit -c lux-ai-2021 -f FILE_NAME -m MESSAGE

• Experiments for Imitation Learning and DQN are conducted with Python 3.7, PyTorch 1.7, CUDA 11.0, kaggle-enviroments 1.8.12, gym 0.19.0.

• Visualization of self-playing replays of PPO and IL are shown in video/ppo_Trim.mp4 and video/IL_Trim.mp4.