System Modeling for Active Noise Control with Reservoir Computing

This paper investigates the use of reservoir computing for active noise control (ANC). It is shown that the ANC problem can be solved by a concatenation of physically present subsystems. These subsystems can be modelled by reservoirs that are trained, using one shot learning. This approach is compared to genetic algorithms tuning a Volterra filter. Experimental results show that our approach works well as system model, meaning that a reservoir trained on white noise performs good on other input signals as well. This is a major advantage over genetic algorithms that generalize rather badly. Furthermore, our approach needs less data and this data can be gathered in one experiment only.