This is a clean and robust Pytorch implementation of Noisy-Duel-DDQN on Atari.

Pong	Enduro

All the experiments are trained with same hyperparameters. Other RL algorithms by Pytorch can be found here.

gymnasium==0.29.1
numpy==1.26.1
pytorch==2.1.0

python==3.11.5

P.S. You can install the Atari environment via pip install gymnasium[atari] gymnasium[accept-rom-license]

python main.py # Train PongNoFrameskip-v4 with DQN

python main.py --Double True --Duel False --Noisy False # Use Double DQN

python main.py --Double False --Duel True --Noisy False # Use Duel DQN

python main.py --Double False --Duel False --Noisy True # Use Noisy DQN

python main.py --Double True --Duel True --Noisy True # Use Double Duel Noisy DQN

If you want to train on different enviroments, just run

python main.py --EnvIdex 20 # Train EnduroNoFrameskip-v4 with DQN

The --EnvIdex can be set to be 1~57, where

1: "Alien",
2: "Amidar",
...
20: "Enduro",
...
57: "Zaxxon"

For more details, please refer to AtariNames.py.

Note that the hyperparameters of this code is a light version (we only use a replay buffer of size 10000 to save memory). Thus, the default hyperparameters may not perform well on all the games. If you want a more robost hyperparameters, please check the DQN paper (Nature).

python main.py --render True --EnvIdex 20 --Double True --Duel True --Noisy False --Loadmodel True --ModelIdex 900 # Play with Enduro

python main.py --render True --EnvIdex 37 --Double True --Duel True --Noisy True --Loadmodel True --ModelIdex 700 # Play with Pong

You can use the tensorboard to record anv visualize the training curve.

• Installation (please make sure PyTorch is installed already):

pip install tensorboard
pip install packaging

• Record (the training curves will be saved at '\runs'):

python main.py --write True

• Visualization:

tensorboard --logdir runs

For more details of Hyperparameter Setting, please check 'main.py'

DQN: Mnih V, Kavukcuoglu K, Silver D, et al. Human-level control through deep reinforcement learning[J]. nature, 2015, 518(7540): 529-533.

Double DQN: Van Hasselt H, Guez A, Silver D. Deep reinforcement learning with double q-learning[C]//Proceedings of the AAAI conference on artificial intelligence. 2016, 30(1).

Duel DQN: Wang, Ziyu, et al. "Dueling network architectures for deep reinforcement learning." International conference on machine learning. PMLR, 2016.

NoisyNet DQN: Fortunato M, Azar M G, Piot B, et al. Noisy networks for exploration[J]. arXiv preprint arXiv:1706.10295, 2017.