Tracking Creative Musical Structure: The Hunt for the Intrinsically Motivated Generative Agent
DOI:
https://doi.org/10.1609/aiide.v9i5.12648Keywords:
music, artificial intelligence, machine learning, adaptive neural network, dynamic bayesian network, reinforcement learningAbstract
Neural networks have been employed to learn, generalize, and generate musical pieces with a constrained notion of creativity. Yet, these computational models typically suffer from an inability to characterize and reproduce long-term dependencies indicative of musical structure. Hierarchical and deep learning models propose to remedy this deficiency, but remain to be adequately proven. We describe and examine a novel dynamic bayesian network model with the goal of learning and reproducing longer-term formal musical structures. Incorporating a computational model of intrinsic motivation and novelty, this hierarchical probabilistic model is able to generate pastiches based on exemplars.