Motion-generation system for violin-playing robot using reinforcement learning differences in bowing parameters due...

Recently, research on human-robot communication attracts many researchers. We believe that music is one of the important channel between human and robot, because it can convey emotional information. In this research, we focus on the violin performance by a robot. Building a system capable of determining performance from a musical score will leads to better understanding communication through music. In this study, we aim to develop a system that can automatically determine bowing parameters, such as bow speed and bowing direction, from musical scores for a violin-playing robot to produce expressive sounds using reinforcement learning. We adopted Q-learning and ε-greedy methods. In addition, we utilized a neural network to approximate the value function. Our system uses a musical score that incorporates the sound pressure value of each note to determine the bowing speed and direction. This study introduces the design of this system. It also presents simulation results on the differences in bowing parameters caused by changes in learning conditions and sound-pressure values. Regarding learning conditions, the learning rate, discount rate, search rate, and the number of units in the hidden layer in the neural network were changed in the simulation. We used the last two bars of the score and the entire four bars in the first phrase of "Go Tell Aunt Rhody." We determined the number of units in each layer and conducted simulations. Additionally, we conducted an analysis by adjusting the target sound pressure for each note in the score. As a result, negative rewards decreased and positive rewards increased. Consequently, even with changes in target sound pressure in both the last two bars and the entire four bars, the violin-playing robot can automatically play from the score by improving reinforcement learning. It has become clear that achieving an expressive performance using this method is possible.