This paper presents Double DQN, a modified version of the Deep Q-Network (DQN) algorithm that reduces overestimation in action values, leading to better performance in reinforcement learning tasks. The original Q-learning algorithm is known to overestimate action values, which can lead to suboptimal policies. The authors show that this overestimation is common in practice and can be mitigated by using Double Q-learning, which separates the selection and evaluation of actions. They propose a specific adaptation of Double Q-learning to DQN, called Double DQN, which improves the accuracy of value estimates and leads to better performance on several games. The paper discusses the problem of overestimation in Q-learning and shows that it can occur due to inaccurate value estimates, regardless of the source of approximation error. The authors demonstrate that overestimation can be reduced by using a target network in DQN, which is used to evaluate the value of the greedy policy. This approach, called Double DQN, leads to more accurate value estimates and better policies. The authors evaluate Double DQN on the Atari 2600 domain, where it outperforms DQN in terms of value accuracy and policy quality. They show that Double DQN achieves state-of-the-art results on the Atari domain, demonstrating that overestimation in DQN was indeed leading to poorer policies. The paper also discusses the robustness of Double DQN to different starting conditions and shows that it performs well even when starting from human trajectories. The authors conclude that Double DQN is a significant improvement over DQN, as it reduces overestimation and leads to more stable and reliable learning. They also show that Double DQN is more robust to challenging evaluation conditions, suggesting that it can find general solutions rather than relying on deterministic sequences of steps. The paper provides a detailed analysis of the performance of Double DQN on a variety of games and shows that it achieves better results than DQN in most cases.This paper presents Double DQN, a modified version of the Deep Q-Network (DQN) algorithm that reduces overestimation in action values, leading to better performance in reinforcement learning tasks. The original Q-learning algorithm is known to overestimate action values, which can lead to suboptimal policies. The authors show that this overestimation is common in practice and can be mitigated by using Double Q-learning, which separates the selection and evaluation of actions. They propose a specific adaptation of Double Q-learning to DQN, called Double DQN, which improves the accuracy of value estimates and leads to better performance on several games. The paper discusses the problem of overestimation in Q-learning and shows that it can occur due to inaccurate value estimates, regardless of the source of approximation error. The authors demonstrate that overestimation can be reduced by using a target network in DQN, which is used to evaluate the value of the greedy policy. This approach, called Double DQN, leads to more accurate value estimates and better policies. The authors evaluate Double DQN on the Atari 2600 domain, where it outperforms DQN in terms of value accuracy and policy quality. They show that Double DQN achieves state-of-the-art results on the Atari domain, demonstrating that overestimation in DQN was indeed leading to poorer policies. The paper also discusses the robustness of Double DQN to different starting conditions and shows that it performs well even when starting from human trajectories. The authors conclude that Double DQN is a significant improvement over DQN, as it reduces overestimation and leads to more stable and reliable learning. They also show that Double DQN is more robust to challenging evaluation conditions, suggesting that it can find general solutions rather than relying on deterministic sequences of steps. The paper provides a detailed analysis of the performance of Double DQN on a variety of games and shows that it achieves better results than DQN in most cases.