Removing additive noise via neuro-fuzzy-based reinforcement learning

Abstract

In this paper, a systematic treatment for developing a noise removal system based on the fundamental principle of reinforcement learning and fuzzy cerebellar model articulation controller (FCMAC) is presented. The proposed system improves its performance over time through two mechanisms. First, the modified stochastic real-valued algorithm, learning from its own mistakes via the reinforcement signal and reinforcing its action to improve future performance, is used for searching the optimal noise spectrum for the overall training system. Second, system states associated with the positive reinforcement are memorized by FCMAC-based neurons, where, in the future, similar states will share the experiences already stored there and then lead the action to a more positive situation. In this work, FCMAC's intrinsically poor approximation of rapidly varying functions is solved by taking the complex semicepstrum. In addition, the FCMAC provides an improvement in accuracy of function approximation without losing the property of generalization, which makes the high fidelity digital signal processing possible.