Musical networks
Introduction to Reinforcement Learning
Introduction to Reinforcement Learning
Learning to Predict by the Methods of Temporal Differences
Machine Learning
| Hi-index | 0.00 |
Reinforcement learning is an exciting possibility for generating music because of its ability to learn without explicit examples and to produce more than one response in a given state. We use reinforcement learning in the second phase of a jazz improvisor that learns to interactively play jazz with a human. The reinforcement signal is based on rules for improvisation. Because of time delays between note played and subsequent reinforcement, a critic adjusts the reinforcement signal. We describe this system and then examine the ability of a temporal difference critic to predict reinforcement for three different sequential musical phenomena. A nonlinear network with a linear TD output unit and context traces on input is able to successfully predict reinforcement values for these sequences and shows promise for use in musical reinforcement learning tasks.