GFANC-RL: Reinforcement Learning-based Generative Fixed-filter Active Noise Control

Abstract

The recent Generative Fixed-filter Active Noise Control (GFANC) method achieves a good trade-off between noise reduction performance and system stability. However, labelling noise data for training the Convolutional Neural Network (CNN) in GFANC is typically resource-consuming. Even worse, labelling errors will degrade the CNN’s filter-generation accuracy. Therefore, this paper proposes a novel Reinforcement Learning-based GFANC (GFANC-RL) approach that omits the labelling process by leveraging the exploring property of Reinforcement Learning (RL). The CNN’s parameters are automatically updated through the interaction between the RL agent and the environment. Moreover, the RL algorithm solves the non-differentiability issue caused by using binary combination weights in GFANC. Simulation results demonstrate the effectiveness and transferability of the GFANC-RL method in handling real-recorded noises across different acoustic paths.²