dueling-DQN-pytorch

very easy implementation of dueling DQN in pytorch

all things are in one file, easily to follow~~

requirement

• tensorflow (for tensorboard logging)
• pytorch (>=1.0, 1.0.1 used in my experiment)
• gym
• ViZDoom

CartPole-v0

for training ddqn in CartPole, just run

python ddqn.py

common, no description~

in CartPole-v0 the network will convergence to 200 episode reward very quickly~~

Visual doom

for training ddqn in Visual doom, just run

python visual_doom.py

for testing ddqn in Visual doom, just run

python visual_doom_test.py

use the basic environment of Visual doom. The agent could play very experienced after about 1000 times of games~~

also provide the dqn training and testing code for comparision.

reference

Dueling Network Architectures for Deep Reinforcement Learning (arxiv)