Learning the evaluation system for Chinese traditional music by evolution strategies

The combination of Computational Intelligence (CI) and music has attracted many efforts. In these studies, how to evaluate artificial music is a hot topic. The design of an evaluation system based on expert experiences is time-consuming and the performance is sometimes not satisfied. Therefore, learning a system from data is a feasible way. However, existing learning-based methods depend highly on the types of data and the downstream machine-learning tasks. In this paper, we model the evaluation system design problem as an offline inverse reinforcement learning problem. Specifically, we use a linear function to act as the evaluation system. The system is trained by natural evolution strategies (NES) using the Chinese traditional folk music data. The experiments show that the learned evaluation system can reasonably evaluate the developed music data and artificial music data generated by existing EC-based mutation operators. Meanwhile, the learned system can not only be applied to Chinese folk music data but also to Chinese opera music data which is not used in the training process.