強化学習によるバイオリン演奏ロボットの動作生成

—割引率と探索率の弓運動への影響—

抄録

We are building a system that can automatically determine the bowing and fingering parameters for our anthropomorphic robot to perform the violin. We adopt the reinforcement leaning with ε-greedy policy, in which a neural network is embedded as a value function. This paper presents simulation results for determining the bow speed and the bowing-direction under the constant bow-force condition. Effects of the discount factor γ and the exploring rate ε on the bowing motion are investigated, while the previous study focused only on the discount factor. Simulation results suggests that by choosing appropriate values for the parameters, we can obtain the bowing parameters with higher reward values than the previous report, which means that the robot can produce sounds close to the target.