Polyphonic music generation generative adversarial network with Markov decision process

In the process of polyphonic music creation, it is important to combine two or more independent melodies through technical treatment. However, due to the diversity of polyphonic music sequences and the limitations of neural networks, it is difficult to create chords or melodies beyond the training data. As the music sequence increases, the probability of the generator producing the same note will increase, which will destroy the coherence of the music. Therefore, this paper proposes a novel polyphonic music creation model, combining the ideas of the Markov decision process (MDP) and Monte Carlo tree search (MCTS) and improving the Wasserstein Generative Adversarial Network (WGAN) theory. Through the zero-sum game and conditional constraints between generator and discriminator, the model in this study is closer to the unconstrained creation of music, and the growth of music sequence will not affect music coherence. Experimental results show that the algorithm proposed here has a better effect on polyphonic music generation than the latest methods.