Systems which evolve music are becoming increasingly popular under the domain of interactive evolution. Steady state neuroevolution can be used for both learning an example and learning through interaction. This provides an avenue to creative composition. However, too much freedom in creation can lead to undesirable results and thus knowledge of music must be incorporated. Furthermore, creation of artworks with patterns requires some sort of memory to store the patterns. We attempt to solve both of these issues through providing a representation for chords based on simple music theory, and using a structure which inherently builds a type of memory. In this paper, we present and perform preliminary evaluations of two distinct, yet related ideas. The first idea is dealing with new implementations for methods of steady state NEAT. The second idea demonstrates the ability of the modified NEAT structure to learn and produce basic chord progressions. 1

Abstract. We present an evolutionary automatic music composer that finds and emphasizes typical patterns in the style of a human composer according to training. First, from a training corpus of melodies the system learns a matrix of conditional probabilities of note transitions; the music is generated as a sequence of notes representing a Markov process with this probability distribution. Then, the probability matrix is iteratively modified by learning from the system’s own output; this emphasizes frequent patterns. After such evolution, the system’s output becomes rich in patterns typical for this composer.