9720508 Franklin, Judy A. Mount Holyoke College POWRE: Positive Human Machine Interaction through Music This award supports an exploration into enabling machines to learn to interact with humans via music. Jazz improvisation is a domain for multi-scale on-line learning systems that requires continuous improvement of interactive performance during learning. Learning takes place at multiple levels and may have poorly defined, changing goal/critic functions. Yet the domain problems have straightforward numerical representations. It is a popular, deep source of human interaction. Reinforcement learning (RL) is well suited. RL is easily used within a numerical context, within a hierarchical structure. A two-level structure is proposed. One RL node creates melody. A high level node learns to "mix" three melody sources: the first RL node, the melody from the human, and a third node that knows about jazz scales. Significantly, RL algorithms can use direct feedback (the improvisation fits rules of jazz) and "indirect" feedback (energy in the human's musical reply increases, human rates it highly, audience claps). The goals are to understand how to obtain reinforcement from a human, to facilitate improving human/machine discourse. Results are relevant to other kinds of discourse. The PI will start a strong research program and continue collaborations with cognitive scientists interested in how humans learn to improvise (in the jazz and more general sense).
