使用深度強化學習技術學習鋼琴技巧

Abstract

現今因為深度學習技巧的出現，使得人類的動作如走路、穿衣服、打籃球等等能夠利用深度學習快速歸納出一個規則。鋼琴是世界上最著名的樂器之一，但是彈鋼琴的技巧即使對人來說也是非常難的一件事。彈奏鋼琴包含閱譜、手指彈奏、指法順序等等困難的技巧，非常難歸納整理成一個原則。因此，使用強化學習技巧學會鋼琴技巧非常具有挑戰性。 此篇論文主旨在於如何將一個已知可以使用深度強化學習方式來訓練的環境一步驟一步驟轉換成使用單隻手掌來彈鋼琴。論文中提出了三項重要的技巧：如何轉換一個可以運行的環境到我們想要的環境；如何將鋼琴的指法轉換成為手指在鋼琴中的位置；如何將鋼琴琴鍵的資訊和手指位置資訊結合讓深度強化學習的網路來學習。 最終結果成功讓手指能夠藉由指法的輔助按下鋼琴上的單音。後續針對結果進行彈黑色琴鍵以及彈白色琴鍵的準確性分析、限制分析以及後續可能性猜測。

Recently the appearance of the skill, deep learning, makes the conclusion of human motions such as walking, wearing clothes and playing basketball come true. Piano is one of the popular instruments, and skills of playing piano is very hard even for a human. It contains reading piano sheet, playing fingers, scheduling fingering and other difficult techniques to play piano, hence these skills are hard to conclude into a rule. In conclusion, using reinforcement learning to learn piano skills can be a contribution to the field of reinforcement learning. The point of this thesis is to transform a workable task using deep reinforcement learning into the task of playing single note using a robot's palm step by step. In this thesis, we show three important methods: how to transform a workable task into our desiring task; how to understand piano fingering and change that into the positions of fingers; how to combine the information of piano keys and information of finger positions for the network agent of deep reinforcement learning to learn. In the end, we successfully make the fingers be able to press down the correct note by the given fingering. We also analyze the accuracy of pressing black piano keys and pressing white piano keys, and generalize the limitations of our methods.According to our work, we conclude some discussion of future works.