Collection of DQN variations from papers, implemented in pytorch
--- Gif of gameplay here
Implementation of Human-level control through deep reinforcement learning
Using a convolutional neural network to take inputs, as well introducing a network termed experience replay, and only periodically updating the Q-value will allow the DQN to perform better than a standard naive DQN
Pong results of DQN trained on 500 episodes
Implementation of Deep Reinforcement Learning with Double Q-learning
Fix DQN's over-estimation of some action values by decoupling the selection from the evaluation of an action
--- results comparison here