In this article, we present a novel domain for user modeling and intelligent agents -agents as specific users' improvisational music companions (IMCs). We also introduce Band-OUT-of-a-Box (BoB), an agent designed to trade live, customized solos with a specific improvising musician/user. A primary technical challenge in this domain is to computationally specify the agent's musical intelligence, a highly personalized and context-dependent concept. To address this issue, we develop an unsupervised machine learning algorithm that can learn to model the musician's behavior after listening to them improvise for awhile. This model learns to abstract bars of solos into user-specific playing modes, which allows the agent to perceive what the musician plays in a user-specific manner. User-specific perception is only part of what the agent needs to respond to the musician. Another key challenge is to "close the learning loop." We address this issue by incorporating the learned model's parameters back into a stochastic process that, when sampled, can produce solos that realize specific user playing modes. By integrating these two components together, BoB can participate in user- and context-specific solo trading. These technologies are evaluated in two simulated real-time contexts, the improvisations of saxophonist Charlie Parker and violinist Stephane Grappelli.
展开▼
