Dueling DQN

E98474

Dueling DQN is a deep reinforcement learning algorithm that separates state-value and advantage estimations within its neural network architecture to improve learning efficiency and stability over standard DQN.

Jump to: Statements Referenced by

Statements (44)

Predicate Object
instanceOf Deep Q-Network variant
deep reinforcement learning algorithm
value-based reinforcement learning method
actionSpaceType discrete
aimsToImprove learning efficiency
training stability
basedOn Q-learning
citationVenue Proceedings of the 33rd International Conference on Machine Learning NERFINISHED
combinesStreamsToEstimate Q-values
commonlyEvaluatedOn Atari 2600 games NERFINISHED
controlType off-policy
domain artificial intelligence
especiallyHelpsWhen many actions have similar value
only a few actions affect the value of the state
extends Deep Q-Network NERFINISHED
field reinforcement learning
hasComponent advantage stream
value stream
hasFullName Dueling Deep Q-Network NERFINISHED
hasKeyIdea decouple representation of state value from representation of advantages for each action
implementedIn DeepMind Atari agent
improvesOver standard DQN NERFINISHED
influenced Rainbow DQN NERFINISHED
introducedBy Hado van Hasselt NERFINISHED
Marc Lanctot NERFINISHED
Matteo Hessel NERFINISHED
Nando de Freitas NERFINISHED
Tom Schaul NERFINISHED
Ziyu Wang NERFINISHED
introducedInPaper Dueling Network Architectures for Deep Reinforcement Learning NERFINISHED
learningParadigm model-free
normalizesAdvantageStream by subtracting mean advantage
oftenCombinedWith Double DQN NERFINISHED
Prioritized Experience Replay NERFINISHED
publishedAtConference ICML 2016 NERFINISHED
separatesEstimationOf advantage function
state-value function
sharesFeatureExtractor between value and advantage streams
usesFunctionApproximator deep neural network
usesLossFunction temporal-difference loss
usesOptimizationMethod Adam optimizer NERFINISHED
stochastic gradient descent
usesTargetNetwork yes
yearIntroduced 2016

Referenced by (2)

Full triples — surface form annotated when it differs from this entity's canonical label.