This paper introduces a new neural network architecture called the dueling network for deep reinforcement learning (DRL). The architecture separates the representation of state values and action advantages, enabling more efficient learning in model-free RL. The dueling network consists of two streams: one for estimating the state value function and another for the action advantage function. These streams share a common convolutional feature learning module and are combined via a special aggregating layer to produce an estimate of the state-action value function Q. This architecture allows the RL agent to outperform state-of-the-art methods on the Atari 2600 domain. The key insight of the dueling architecture is that it can learn which states are valuable without needing to estimate the effect of each action for every state. This is particularly useful in states where actions do not significantly affect the environment. The architecture is implemented as a single Q-network with two streams that replace the popular single-stream Q-network in existing algorithms. The dueling network automatically produces separate estimates of the state value function and advantage function without any extra supervision. The paper evaluates the performance of the dueling architecture on a policy evaluation task and on general Atari game-playing. In the policy evaluation task, the dueling architecture outperforms the traditional Q-network, especially as the number of actions increases. In the Atari game-playing experiments, the dueling architecture significantly improves performance over single-stream Q-networks, achieving human-level performance on 42 out of 57 games. The results show that the dueling architecture is complementary to algorithmic innovations and can be combined with other improvements such as prioritized experience replay to further enhance performance. The dueling architecture is also shown to be robust to different starting conditions and to perform well in a wide range of Atari games. The results demonstrate that the dueling architecture leads to substantial improvements in performance in deep reinforcement learning, particularly in the challenging Atari domain. The architecture is implemented with a shared convolutional feature learning module and is compatible with existing and future RL algorithms. The paper concludes that the dueling architecture is a significant advancement in deep reinforcement learning, offering a more efficient and effective approach to model-free RL.This paper introduces a new neural network architecture called the dueling network for deep reinforcement learning (DRL). The architecture separates the representation of state values and action advantages, enabling more efficient learning in model-free RL. The dueling network consists of two streams: one for estimating the state value function and another for the action advantage function. These streams share a common convolutional feature learning module and are combined via a special aggregating layer to produce an estimate of the state-action value function Q. This architecture allows the RL agent to outperform state-of-the-art methods on the Atari 2600 domain. The key insight of the dueling architecture is that it can learn which states are valuable without needing to estimate the effect of each action for every state. This is particularly useful in states where actions do not significantly affect the environment. The architecture is implemented as a single Q-network with two streams that replace the popular single-stream Q-network in existing algorithms. The dueling network automatically produces separate estimates of the state value function and advantage function without any extra supervision. The paper evaluates the performance of the dueling architecture on a policy evaluation task and on general Atari game-playing. In the policy evaluation task, the dueling architecture outperforms the traditional Q-network, especially as the number of actions increases. In the Atari game-playing experiments, the dueling architecture significantly improves performance over single-stream Q-networks, achieving human-level performance on 42 out of 57 games. The results show that the dueling architecture is complementary to algorithmic innovations and can be combined with other improvements such as prioritized experience replay to further enhance performance. The dueling architecture is also shown to be robust to different starting conditions and to perform well in a wide range of Atari games. The results demonstrate that the dueling architecture leads to substantial improvements in performance in deep reinforcement learning, particularly in the challenging Atari domain. The architecture is implemented with a shared convolutional feature learning module and is compatible with existing and future RL algorithms. The paper concludes that the dueling architecture is a significant advancement in deep reinforcement learning, offering a more efficient and effective approach to model-free RL.