Determined Blind Source Separation Combining Independent Low-rank Matrix Analysis with Optimized Parameters and Q-learning

This study utilizes Q-learning to dynamically change the optimized parameters of independent low-rank matrix analysis (ILRMA). Notably, ILRMA, which combines independent vector analysis and non-negative matrix factorization, is a novel methodology adopted to realize multichannel blind source separation (BSS). In previous studies, ILRMA has used optimized parameters to improve separation efficiency. In this study, however, an optimized parameter is obtained from the parametric majorization–equalization algorithm, which adjusts the convergence speed to avoid poor local solutions. Two other optimized parameters are obtained using the isotropic complex Student’s t-distribution, which adjusts the probability distribution to conform to the target mixed audio source distribution. To further improve the performance of BSS, this paper proposes a Q-learning off-policy temporal difference control algorithm (reinforcement learning) to dynamically change the three optimized parameters. To ensure the simplicity and efficiency of Q-learning, n-armed bandits are used to replace traditional high-dimensional Q-tables. Furthermore, experiments are conducted using instruments and vocal multichannel BSS tasks. The results confirm the validity of the proposed method.