基于强化学习Actor-Critic算法的音乐生成

MUSIC GENERATION BASED ON REINFORCEMENT LEARNING ACTOR-CRITIC

提出一种利用强化学习Actor-Critic(A-C)训练神经网络生成音乐的方法。常规的LSTM音乐生成网络在生成音乐时并没有考虑到实际的作曲情况,只是通过先前训练保存的策略来选择下一个音符,所以生成的音乐稳定性差、风格模糊。引入一个经过训练的Critic网络,该网络能够评估LSTM网络输出音符的价值,以此更新LSTM网络的生成策略。这形成了一个更接近生成阶段的训练过程,并允许优化特定的音乐风格,所以生成的音乐结构稳定,更具风格。对该方法生成的音乐进行验证,证明了其有效性。

This paper proposes a music generation method using reinforced learning Actor-Critic(A-C)to train neural network.The LSTM music generation network does not take the actual composition into account when generating music.It only selects the next note by the strategy saved by the previous training,so the generated music is poor in stability and style.This paper introduced a trained Critic network.It could evaluate the value of the output notes of the LSTM network,so as to update the generation strategy of the LSTM network.It formed a training process closer to the generation phase and allowed for the optimization of specific musical styles.Therefore,the generated music is stable in structure and more stylistic.The music generated by this method is verified,and the result proves its effectiveness.