Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 

Vectorized value-based reinforcement learning

Reinforcement learning

DQN/ and DoubleDQN/ implement the same environment set with batched environment wrappers, replay buffers, Q-networks, agents, configurations, and training entrypoints.

Algorithm Target TensorFlow implementation
DQN (r + \gamma\max_a Q_{\bar\theta}(s',a)) Tensor batches, replay sampling, target networks, GradientTape, tf.function
Double DQN (r + \gamma Q_{\bar\theta}(s',\arg\max_a Q_\theta(s',a))) Separate online/target action roles in the agent update

CartPole and Acrobot include test scripts. Training files also contain TensorBoard summary code. No training score is documented here because the repository does not retain a benchmark report as source.