Active noise control (ANC) has been intensively studied for decades. The most classical ANC algorithm should be the filtered-x least mean square (FxLMS) algorithm, which needs the model of the secondary path to work. Thus, the residual error of the ANC system is closely related to the preciseness of the secondary path model. In many applications, the secondary path is often time-varying. Therefore, off-line identification of the secondary path is not applicable. However, on-line identification often requires an additional white noise as a stimulating signal of the secondary path, which will deteriorate the final noise reduction effect. This paper proposes an improved artificial bee colony (ABC) algorithm for ANC system, which does not require identification of the secondary path. In order to guarantee the convergence of the algorithm and accelerate the convergence speed, this paper introduces a variable forgetting factor into the fitness function, and improves the traditional ABC algorithm by integrating LMS algorithm into the ABC algorithm. A single channel ANC system equipped with an FPGA hardware platform is set up in an anechoic chamber, and experiments show that the proposed algorithm can produce a satisfactory noise reduction effect without modeling the secondary path.