Symmetry-Based Data Augmentation Improves DDPG for Aircraft Control
A research article available on arXiv presents a novel symmetric data augmentation technique aimed at enhancing sample efficiency in offline reinforcement learning, specifically for lateral attitude tracking control in fixed-wing aircraft using Deep Deterministic Policy Gradient (DDPG). This approach leverages the symmetry present in Markov Decision Processes (MDPs) to create augmented samples, thereby broadening the coverage of the state-action space. Additionally, a dual-critic framework is introduced, where a second critic is trained on these augmented samples to further optimize sample usage. The symmetry of the aircraft model is confirmed, and simulations indicate that using augmented samples leads to faster policy convergence. This paper falls under the category of Computer Science > Machine Learning and was submitted on July 15, 2024.
Key facts
- Paper proposes symmetric data augmentation for DDPG in offline RL.
- Method exploits symmetry of dynamical systems for state-transition prediction.
- Augmented samples enhance coverage rate of state-action space.
- Dual-critic structure introduced: second critic trained on augmented samples.
- Aircraft model verified to be symmetric.
- Flight control simulations demonstrate accelerated policy convergence.
- Paper submitted on July 15, 2024.
- Categorized under Computer Science > Machine Learning.
Entities
Institutions
- arXiv