This is a course project of the 2048 Game for PKU's graduate students' lecture Reinforcement Learning in Autumn 2021.
Python >= 3.7 PyTorch >= 1.6
Please pay attention that Python versions earlier than 3.7 may not support this project. F-string is widely used in this project for performance. However, earlier versions of Python do not support f-string.
We implement three network architectures and two encoding methods. For details, please read model/Layers.py
This project involves many carefully-designed hyperparameters defined in train_AI.py.
We use the default hyperparameters in train_AI.py for training (20w episodes)
| Avg_score | 4096 rate | 2048 rate | 1024 rate | 512 rate | 256 rate | |
|---|---|---|---|---|---|---|
| emd + Conv2 | 14497.81 | 0.00% | 11.85% | 71.35% | 96.40% | 99.40% |
| emd + CNN | 13734.06 | 0.00% | 11.35% | 71.20% | 96.80% | 99.80% |
| emd + CNNpool | 11126.82 | 0.00% | 3.85% | 50.00% | 91.70% | 99.70% |