Research And Implementation Of Music Generation Based On Deep Reinforcement Learning

Music has always occupied an important position in the daily life of mankind.With the development of society,the demand for music is increasing.The traditional music creation method has a long cycle and requires the creator to have professional music knowledge.Therefore,the use of computer intelligence to generate music has become a hot research topic.Although the use of deep learning avoids the manual construction of features,the existing music generation algorithms still have many shortcomings.For example,the difference between the training phase and the generation phase is not taken into account,and this difference will increase the error,thereby affecting the quality of the generated music.Just let the network learn the dependency relationship between the notes from the music data,without considering the real music creation process,resulting in poor harmony of the generated music.Therefore,this article aims at the above problems,in order to obtain higher quality and more artistic aesthetic music,research on the intelligent generation of music.The main contributions of this article are as follows:1.In order to complete the music generation task,this thesis builds a characterlevel music generation model Melody＿LSTM based on Long Short-Term Memory(LSTM).In this thesis,music is converted into the form of note sequences for processing,and encoding techniques are used to convert the note sequences into the form of note vectors and input into the neural network.By taking the input note sequence at the next moment as the target value,a training process of supervised learning is formed,so that the Melody＿LSTM network can learn the dependency relationship between notes on the time scale.Through the generation experiment,it is proved that the Melody＿LSTM network can learn the dependency relationship between the notes from the music data,and can complete the end-to-end music generation task.2.In view of the problems in the existing music generation model,in order to allow the model to take into account the impact of chord progression and music theory rules on the selection of notes,this thesis builds an ACMG music generation model based on the Melody＿LSTM network and uses deep reinforcement learning Actor-Critic algorithm.The model introduces a Critic network,which consists of a chord reward mechanism,a music theory reward mechanism,and a state value reward mechanism.The Critic network can evaluate the value of the generated notes at the current moment.The music generation network further updates its own generation strategy according to its value,which forms a training process closer to the generation stage and allows the optimization of specific music styles.Finally,it is proved by comparative experiments the validity of the ACMG model.3.In order to obtain high-quality music,based on the existing open source music databases Lakh MIDI and NSynth,this thesis uses deep learning technology to build a clean and regular artificial simulation music database(ASMD),and makes different The comparison experiment between the databases proves that the database has a positive effect on the training of deep models.