- 學位論文
透過深度學習自動生成太鼓達人遊戲譜面
Automatic Generation of Taiko Games Using Deep Learning Methods
摘要
太鼓達人是世界上流行的節奏遊戲之一，它模擬了音樂下的太鼓演奏模式。然而，在這個遊戲之中存在一個煩人的系統，就是節奏點的判斷與安排。而這件事往往需要耗費大量的人力跟時間去完成，才得以讓遊戲譜面跟歌曲形成一致的遊玩模式，而本篇論文則是研發可以自動產生遊戲譜面的方式。 本篇論文設計了兩個步驟，第一是時間點的生成步驟，它決定了何時應該放置節拍。第二是節奏點類型的生成步驟，它決定了第一步驟的時間點上所生成的節奏類型。 在這項研究中，有使用一些特殊的方法。首先是模糊標籤(fuzzy label)的使用，因為訓練數據極度不平衡，模糊標籤可以使其變得更平滑，也可以提高數據的含標籤量，促使數據變得足夠平衡。再來是讓數據的移動幅度與快速傅立葉轉換的窗口大小相符合，如此一來，就不會出現數據在前處理的過程中被分成兩份。最後是使用數據量更大的卷積長短期記憶(C-LSTM)模型。 在測試數據上面，我們將預測時間點步驟的F-score，在六十四分之一拍的前提下，從0.7769提高到0.8312，在三十二分之一拍的前提下，從0.8765提高到0.9068。
並列摘要
Taiko no Tatsujin is one of the popular rhythm games in the world. It simulates playing the taiko drum in time with music. However, this game exists a system struggle, that is to replicate human-like patterning. The placement of game objects in relation to each other to form congruent patterns based on events in the song. This thesis introduces two steps to generate the beatmaps. First step is timestamp generation, it decides when to put the beat. Second step is action type generation, it decides what to put on the time in first step. In this research, some special methods are purposed. First is “fuzzy label”, since the training data is extreme unbalance. Fuzzy label can make the changing smoother and increase the positive rate of data, which make the data balance enough to train. Second is make the stride fit the window lengths of the music. Which the training data will not be cut into two part by unfit stride. Third is the C-LSTM model with larger data size. On the test data, the thesis improved the F-score of timestamp prediction from 0.7769 to 0.8312 for 64 semi-notes and 0.8765 to 0.9068 for 32 semi-notes.