Fourier Features in Reinforcement Learning with Neural Networks
Abstract
In classic Reinforcement Learning (RL), encoding the inputs with a Fourier feature mapping is a standard way to facilitate generalization and add prior domain knowledge. In Deep RL, such input encodings are less common since they could, in principle, be learned by the network and may therefore seem less beneficial. In this paper, we present experiments on Multilayer Perceptrons (MLP) that indicate that even in Deep RL, Fourier features can lead to significant performance gains in both rewards and sample efficiency. Furthermore, we observe that they increase the robustness with respect to hyperparameters, lead to smoother policies, and benefit the training process by reducing learning interference, encouraging sparsity, and increasing the expressiveness of the learned features. However, a major bottleneck with conventional Fourier features is that the number of features increases exponentially with the state dimension. As a remedy, we propose a simple, light version that only has a linear number of features yet empirically provides similar benefits. Our experiments cover both shallow/deep, discrete/continuous, and on/off-policy RL settings.
Origin | Files produced by the author(s) |
---|