Implementation of the original Deep Q-network algorithm from Mnih et al. 2015.
Python 3.6+ and TensorFlow are required.
pip3 install git+git@github.com:JuliusKunze/nevermind.git
will install nevermind together with minimal requirements.
Solve the cartpole gym environments in a few minutes on a CPU:
from nevermind import configurations
configurations.train_cartpole()Train for the lunar lander environment - typically solved in a few hours on a CPU:
configurations.train_lunar_lander()Training for the Atari 2600 game breakout takes days on a GPU:
configurations.train_atari(game='breakout')Models and training summary plots (see below) are saved to the directory ./data by default.
Learned value and advantage function for the cartpole task:
Cartpole training summary of 10 runs:
Summary of a successful lunar lander training run:
10M timesteps of training for the breakout Atari environment:
