Abstract: We propose a frequency-domain state representation to improve the performance and reduce the computation and data requirements of reinforcement learning. This approach is tailored to ...