This repo is a simple test code for "Episodic Memory Deep Q-Networks"
by Zichuan Lin, Tianqi Zhao, Guangwen Yang and Lintao Zhang.
And their EMDQN source code can be gotten here using Tensorflow.
This code is a simple version using PyTorch.
The users can modify the code to suit your own testing environments (DISCREATE ACTION SPACE).
Episodic Memory DQN
- Python >= 3.6.0 (optional)
- PyTorch == 1.7.0 (optional)
- OpenAI Gym[Atari]
- Scikit-Learn == 1.0.2 (optional)
- To run this code,
cdinto the root directory and run :python main.py --env PongNoFrameskip-v4 - The kernel updating codes for EMDQN algorithm:
./core/emdqn.py - The Episodic Memory using LRU_KNN:
./utils/lru_knn.py - The off-policy replay buffer:
./components/replay_buffer.py - The networks include base Q network and Dueling Q network:
./components/networks.py - The Hyper-parameters can be modified in:
./components/arguments.pyThe details can be seen in the original paper.
- DQN, DuelingDQN, EMDQN
- Save and Load Pre-Trained Models.
- CUDA supported.
- Modify and Suit Atari Games.
- Modify Prioritized ReplayBuffer.