Twin Delayed Deep Deterministic Policy Gradient (TD3) - kairproject/kair_algorithms_draft GitHub Wiki

TD3 (Fujimoto et al., 2018) is an extension of DDPG (Lillicrap et al., 2015), a deterministic policy gradient algorithm that uses deep neural networks for function approximation. Inspired by Deep Q-Networks (Mnih et al., 2015), DDPG uses experience replay and target network to improve stability. TD3 further improves DDPG by adding clipped double Q-learning (Van Hasselt, 2010) to mitigate overestimation bias (Thrun & Schwartz, 1993) and delaying policy updates to address variance.

• Example Script on LunarLander
• ArXiv Preprint