NASDAQ: Normalized Observation Space Dynamics-Augmented Q-Learning

The paper introduces Normalized Observation Space Dynamics-Augmented Q-learning (NASDAQ), a novel framework designed to enhance model-free reinforcement learning by addressing the challenge of unbalanced reconstruction losses in low-dimensional observations. NASDAQ employs a normalization method that balances losses and gradients across observation dimensions, facilitating improved dynamics prediction for both low- and high-dimensional inputs. Experimental results indicate that NASDAQ outperforms existing model-based and self-predictive RL methods while significantly reducing training time, making it a valuable approach for practitioners focused on improving sample efficiency and performance in diverse RL tasks.